# BCI Report — full text > Every English page of https://bci.report as Markdown, in the order of https://bci.report/llms.txt. # EEG results you can check. Open EEG evaluation 8 fixed protocols on 7 public datasets carry the core matrix’s 14 decoding methods and, beside them, 14 foundation encoders as frozen probes. Questions beyond the matrix cover transfer, calibration and model adaptation, reliability, clinical groups and data quality, signals other than scalp EEG, and small pilot studies. The open register lists every EEG dataset and model we know of, with how far each has been checked. Core matrix - Protocols: 8 - Datasets: 7 - Comparisons: 69 - Methods: 14 [Also on these protocols: 14 foundation encoders, frozen →](https://bci.report/protocols/) Across the site: [17 questions](https://bci.report/topics/) · [32 datasets with results](https://bci.report/datasets/) · [22 methods with result pages](https://bci.report/methods/) Core benchmark matrix ## 14 methods × 8 protocols Balanced accuracy in percent, and command detection on idle. Each method’s score on each protocol; each column heading gives the cohort size and chance level. | Method | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) — n=10 · chance 50% | [Idle & command](https://bci.report/protocols/idle/) — n=4 · detection | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) — n=70 · chance 2.5% | [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) — n=70 · chance 2.5% | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) — n=36 · chance 50% | [P300 target ERP](https://bci.report/protocols/p300-target/) — n=21 · chance 50% | [Semantic target ERP](https://bci.report/protocols/semantic-target/) — n=30 · chance 50% | [Sleep staging](https://bci.report/protocols/sleep-scalp/) — n=20 · chance 20% | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model · 8 of 8 | 62.9 | 18.3 | 33.7 | 27.6 | 62.3 | 55.0 | 55.8 | 71.3 | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · 8 of 8 | 68.6 | 3.3 | 52.8 | 38.7 | 61.3 | 50.8 | 57.4 | 45.2 | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model · 8 of 8 | 71.0 | 1.7 | 55.8 | 44.1 | 67.6 | 53.3 | 61.4 | 56.5 | | [LaBraM](https://bci.report/methods/labram/) — Foundation model · 8 of 8 | 53.5 | 18.3 | 10.8 | 12.9 | 64.6 | 49.4 ≤ chance | 53.2 | 71.1 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · 8 of 8 | 68.8 | 43.3 | 23.0 | 16.2 | 67.9 | 49.7 ≤ chance | 56.7 | 55.0 | | Spectral ridge — Classical method · 7 of 8 | 54.9 | | 50.7 | 48.2 | 56.8 | 51.0 | 51.0 | 72.2 | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method · 6 of 8 | 47.1 ≤ chance | 23.3 | – | – | 68.3 | 50.0 ≤ chance | 52.7 | 50.1 | | [MDM](https://bci.report/methods/riemannian/) — Classical method · 3 of 8 | 53.8 | 35.0 | | | 71.5 | | | | | [TS+LR](https://bci.report/methods/riemannian/) — Classical method · 3 of 8 | 58.9 | 35.0 | | | 65.5 | | | | | [FBCCA](https://bci.report/methods/fbcca/) — Classical method · 2 of 8 | – | – | 79.1 | 72.3 | – | – | – | – | | [Standard CCA](https://bci.report/methods/cca/) — Classical method · 2 of 8 | – | – | 63.1 | 57.6 | – | – | – | – | | Temporal ridge — Classical method · 2 of 8 | – | – | | | – | 53.8 | 56.3 | – | | [xDAWN+LDA](https://bci.report/methods/xdawn/) — Classical method · 2 of 8 | – | – | | | – | 60.4 | 56.8 | – | | [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/) — Classical method · 2 of 8 | – | – | | | – | 61.9 | 60.3 | – | Bar = position between that protocol's chance level and 100% Not applicable to this protocol Not run [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [`core-matrix-20261106.json` ↓](https://bci.report/data/core-matrix-20261106.json) Bars are comparable **down** a column and not **across** one: chance level is 50% for the two-class tasks, 20% for sleep staging and 2.5% for 40-class SSVEP, and every protocol uses a different cohort and electrode layout. There is no overall score. Read the idle column with the false-activation rate in its protocol. 13 further foundation encoders, run as frozen probes on the same eight protocols (BrainOmni Base on six of them). They are checkpoints of 11 models; 3 more checkpoints of one of them form a masking ablation, 16 in all. They sit outside the matrix. [The foundation-encoder rows on the first protocol →](https://bci.report/protocols/mi-rest/#foundation-v9) Beside them sits EEGPT, from the same runs as the published LaBraM rows and, except on idle, the CBraMod rows, with only the encoder swapped. [EEGPT’s row on the first protocol →](https://bci.report/protocols/mi-rest/#eegpt) [Jev-style questions on EEG: the evidence](https://bci.report/jev-style/): Decision research · “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. Three routes, each answered by this site’s own runs on public EEG data. · Read the evidence → One protocol at a time ## The evidence behind each score ## Rest versus right-hand imagery ds003810 · 10 subjects · 1,200 epochs · 5 participant-disjoint folds · 15 channels · 2-second windows Measured here ### Results at a glance [CSV ↓](https://bci.report/data/mi-rest-results.csv) | Model / configuration | Balanced accuracy | Macro F1 | Compute time | | --- | --- | --- | --- | | Spectral ridge — Supervised fit · 15 ch | **54.9%** — Descriptive 95% interval: 53.2–56.7% | **0.544** — Mean across held-out participants | 0.2s | | CSP+LDA — Supervised fit · 15 ch | **47.1%** — Descriptive 95% interval: 43.8–50.2% At or below the 50% chance level | **0.406** — Mean across held-out participants | 8.8s | | LaBraM — Frozen encoder + ridge head · 15 ch | **53.5%** — Descriptive 95% interval: 51.4–55.4% | **0.509** — Mean across held-out participants | 2.2s | | CBraMod — Frozen encoder + ridge head · 15 ch | **62.9%** — Descriptive 95% interval: 59.6–66.5% | **0.617** — Mean across held-out participants | 5.0s | | EEGNet — Scratch · 20 epochs · 15 ch | **71.0%** — Descriptive 95% interval: 65.0–76.3% Single seed — the highest of 3 runs (68.17–71.00%, mean 69.47%) | **0.680** — Mean across held-out participants | 55.5s | | TS+LR — Supervised fit · 15 ch | **58.9%** — Descriptive 95% interval: 54.9–63.9% | **0.534** — Mean across held-out participants | — | | MDM — Supervised fit · 15 ch | **53.8%** — Descriptive 95% interval: 50.8–57.9% | **0.420** — Mean across held-out participants | — | | ShallowFBCSPNet — Scratch · 20 epochs · 15 ch | **68.8%** — Descriptive 95% interval: 63.4–73.9% | **0.672** — Mean across held-out participants | — | | Deep4Net — Scratch · 20 epochs · 15 ch | **68.6%** — Descriptive 95% interval: 63.3–74.0% | **0.659** — Mean across held-out participants | — | | 13 further foundation encoders, frozen [Every row with its footnote →](https://bci.report/protocols/mi-rest/#foundation-v9) | | | | | BrainOmni Base — Frozen encoder + ridge head · 15 ch | **58.8%** — Descriptive 95% interval: 57.0–60.8% | **0.582** — Mean across held-out participants | — not published | | CodeBrain — Frozen encoder + ridge head · 15 ch | **58.8%** — Descriptive 95% interval: 54.5–62.6% | **0.573** — Mean across held-out participants | — not published | | eeg-fm-masking MAE (r=9 cm, L=2) — Frozen encoder + ridge head · 15 ch | **68.1%** — Descriptive 95% interval: 64.3–72.3% | **0.677** — Mean across held-out participants | — not published | | EEGMamba — Frozen encoder + ridge head · 15 ch | **67.3%** — Descriptive 95% interval: 63.6–70.6% | **0.669** — Mean across held-out participants | — not published | | ERP-FM Base — Frozen encoder + ridge head · 15 ch | **54.8%** — Descriptive 95% interval: 51.8–57.8% | **0.524** — Mean across held-out participants | — not published | | LUNA Base — Frozen encoder + ridge head · 15 ch | **52.6%** — Descriptive 95% interval: 50.2–54.9% | **0.503** — Mean across held-out participants | — not published | | LUNA Large — Frozen encoder + ridge head · 15 ch | **49.9%** — Descriptive 95% interval: 46.6–53.5% At or below the 50% chance level | **0.486** — Mean across held-out participants | — not published | | REVE Base — Frozen encoder + ridge head · 15 ch | **66.2%** — Descriptive 95% interval: 61.8–70.6% | **0.640** — Mean across held-out participants | — not published | | REVE Large — Frozen encoder + ridge head · 15 ch | **70.2%** — Descriptive 95% interval: 64.0–76.1% | **0.673** — Mean across held-out participants | — not published | | SingLEM — Frozen encoder + ridge head · 15 ch | **72.0%** — Descriptive 95% interval: 66.6–76.6% | **0.715** — Mean across held-out participants | — not published | | ST-EEGFormer Base — Frozen encoder + ridge head · 15 ch | **58.1%** — Descriptive 95% interval: 53.7–63.3% | **0.541** — Mean across held-out participants | — not published | | ST-EEGFormer Large — Frozen encoder + ridge head · 15 ch | **62.4%** — Descriptive 95% interval: 57.4–67.8% | **0.586** — Mean across held-out participants | — not published | | ZUNA 1.1 — Frozen encoder + ridge head · 15 ch — Model card: research use only, not for diagnosis or clinical use. | **50.7%** — Descriptive 95% interval: 49.2–52.1% Interval reaches the 50% chance level Exposure unknown | **0.499** — Mean across held-out participants | — not published | | EEGPT, frozen · the same runs as the published LaBraM and CBraMod rows [Its row with its footnotes →](https://bci.report/protocols/mi-rest/#eegpt) | | | | | EEGPT — Frozen encoder + ridge head · 15 ch | **58.3%** — Descriptive 95% interval: 54.9–61.8% | **0.555** — Mean across held-out participants | — not published | Scroll the table sideways for the remaining columns. [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-mi-rest.csv) · [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/mi-rest/#eegpt-terms) Masking ablation: three sibling checkpoints of eeg-fm-masking Read against the paper-recommended checkpoint, the matrix row above: the same architecture, corpus and recipe, with another masking framework or geometry. | eeg-fm-masking MAE (r=9 cm, L=2) — Frozen encoder + ridge head · 15 ch — matrix row above | **68.1%** — Descriptive 95% interval: 64.3–72.3% | **0.677** — Mean across held-out participants | — not published | | --- | --- | --- | --- | | eeg-fm-masking JEPA (r=9 cm, L=2) — Frozen encoder + ridge head · 15 ch | **63.2%** — Descriptive 95% interval: 58.2–67.9% | **0.620** — Mean across held-out participants | — not published | | eeg-fm-masking JEPA (r=one channel, L=1) — Frozen encoder + ridge head · 15 ch | **63.3%** — Descriptive 95% interval: 59.0–67.4% | **0.628** — Mean across held-out participants | — not published | | eeg-fm-masking MAE (r=one channel, L=1) — Frozen encoder + ridge head · 15 ch | **64.9%** — Descriptive 95% interval: 60.3–69.4% | **0.645** — Mean across held-out participants | — not published | Compute time covers fitting and prediction, possibly with accelerator waiting, and excludes data preparation. Visual comparison ### Accuracy by configuration CC0-1.0 · Aggregate research results [Original dataset ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) [Use & privacy →](https://bci.report/data-use/) [Report an issue with this result →](https://bci.report/data-use/#corrections) **Limits** Ten-person laboratory task with prompted rest. This is not continuous-idle monitoring or a physical low-channel headset test. Questions ## What the evidence can answer Figures under one question often share participants: they are repeated conditions within its protocols. ### Transfer Does a decoder still work after a change of sensor, display, electrode layout, body movement or session? - [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/): Sensor transfer · What changes when a decoder crosses between two native eight-channel recordings from the same 102 people? · 2 s SSVEP · 12 targets · balanced accuracy - [Screen to VR](https://bci.report/topics/screen-to-vr/): Context transfer · The same 21 people calibrated on a PC screen and tested in a VR headset, and the reverse: the level reached after the switch, with no same-display score to compare. Plus: a P300 decoder trained at one image rate and tested at another. · P300 · 21 people in VR · 9 people across image rates - [Fewer electrodes](https://bci.report/topics/fewer-electrodes/): Montage · In-ear against scalp for sleep, four posterior electrodes against sixteen for eyes open or closed, foundation encoders on BETA with eight and four electrodes, and a simulated cap shift of one electrode spacing on BETA: channel sets from specific research recordings. · Two paired comparisons · 10 and 19 people · BETA, 8 and 4 electrodes, 70 people · a simulated cap shift - [On the move](https://bci.report/topics/on-the-move/): Motion robustness · Standing, walking and running results, with scalp and ear recordings and incompatible time windows kept apart. · SSVEP balanced accuracy · ERP ROC AUC - [Later sessions](https://bci.report/topics/later-sessions/): Session transfer · same person · Trained on a person’s earlier session and tested on a later one, with no labels from it: motor imagery across recording sessions, with simple baselines and frozen CBraMod; target detection at the publisher’s nominal later visits; and continuous cursor tracking, where a constant baseline does better. · Motor imagery, WBCIC-SHU, 51 and 11 people · RSVP, 15 people · cursor tracking, a negative result ### Adapting models How much calibration a decoder needs, which part of a pretrained model to update, whether pretraining helps, whether one model can answer several questions, including questions asked in words, and what lets a decoder answer an SSVEP frequency it never trained on. - [How much calibration?](https://bci.report/topics/calibration-budget/): Calibration budget · Twelve, 24 or 48 labeled trials on a wearable SSVEP task, and up to 40 from a person’s next motor-imagery session: the gain differs by method, and some people’s scores fall. · SSVEP · 102 people · motor imagery, next session · 51 people - [Which part to update?](https://bci.report/topics/model-adaptation/): Model adaptation · LaBraM on new people, same task, zero labels from the test person: head only, last block or LoRA, printed beside the core matrix’s frozen readout. Plus: the next-day experiment, run and held. · EEGMAT, 36 people · three update rules · next day: held - [Does pretraining help?](https://bci.report/topics/does-pretraining-help/): Representation controls · Matched pretrained and constructor-random encoders under fixed and train-selected readout settings, and the same control for five further checkpoints on three tasks. Plus further foundation models in the same frozen recipe: three checkpoints and, narrowly, EEGPT above every published row on sleep staging, none above training-free CCA on BETA. · Three tasks · eight encoders · three random initializations · 23 further checkpoints - [One model, several questions](https://bci.report/topics/shared-encoder/): Shared encoder · route 2 · One shared encoder with a fixed output per question, against a head told which question it is answering and against a separate model per question, at matched data and compute — with what each costs. · Motor imagery, OpenBMI, 51 people · sleep, BOAS · EESM19, crude - [Questions in language](https://bci.report/topics/questions-in-language/): Questions in language · route 3 · One small head over frozen EEG features, asked each question by a number, a label template or a description: it works only for questions it was trained on, and not on every dataset. · SSVEP, BETA, 70 people · sleep, BOAS · secondary: EESM19, OpenBMI, Wearable-102 - [New SSVEP frequencies](https://bci.report/topics/new-ssvep-frequencies/): Route 3 follow-ups · Fourteen follow-ups: thirteen on route 3’s BETA cohort, five on Wearable-102, one of them only there. Heads that read reference spectra or stimulus phase stayed far below CCA, which needs no training; a small learned correction on top of CCA added a few points on BETA and changed little on Wearable-102; filter-bank CCA, untrained too, scored above CCA in every setting tested, and an untrained harmonic sum came out nearly level with a learned eigenvalue head. · SSVEP, BETA, 70 people · 8 held-out frequencies · Wearable-102, 102 people ### Reliability & clinical When a decoder should not act, what accuracy hides in sleep staging, what resting-state EEG says about a clinical group, and what a public dataset’s files show when checked end to end. - [When not to act](https://bci.report/topics/when-not-to-act/): Abstention · decision research · Command detection read with its false-activation rate; when a decoder should decline to decide, on two protocols; and the decision-research plan, with its three routes. · Idle, 4-person pilot · non-control, 20 people · reliable decisions, 36 and 70 people - [Sleep-stage balance](https://bci.report/topics/sleep-staging/): Sleep staging · simple baselines · Two simple baselines in 25 healthy sleepers and 55 people with sleep apnoea, as two separate experiments: accuracy beside balanced accuracy, and the stage the stager never predicts. · Dreem DOD-H and DOD-O · 25 and 55 nights · balanced accuracy - [Clinical groups](https://bci.report/topics/clinical-groups/): Clinical research · A 149-person Parkinson's and control comparison beside an age-and-sex-only comparator; neither is a diagnosis. · 149 people · one site · balanced accuracy - [Checked end to end](https://bci.report/topics/data-quality/): Data quality and reproducibility · Public EEG datasets read from their own files: Shin2022’s runs, duplicates and valid windows counted before a participant-held-out check, where the EEG arm sits close to a control that reads only trial duration; training-free CCA on Kim beta-range SSVEP and on every eldBETA participant, with filter-bank CCA beside it there; MIND’s stored markers, read on record IDs; two one-person cases, Damm and NeBULA, with counts and no score; and a timing hold on one converted ZuCo run. · Shin2022, 10 people · Kim beta-range · MIND · eldBETA, 100 people · one-person cases, counts only · ZuCo, status only ### Other signals & pilots What simple baselines show on signals other than scalp EEG, in people and in one macaque, and what small pilot studies show. - [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/): Other signals · human and macaque · fNIRS telling right-hand, left-hand and foot tapping apart in people the classifier never saw; finger velocity from one macaque’s motor-cortex spiking, where a ridge decoder lowers the squared error but not the absolute error; and a numerical coverage check of fNIRS recordings. Intracortical handwriting and functional ultrasound: status only. · fNIRS, 30 people · intracortical spiking, 1 macaque · fNIRS coverage, 12 people - [Pilot studies](https://bci.report/topics/pilot-studies/): Pilot studies · small cohorts · Small pilots: injected orthosis errors in one cohort, including a frozen EEG foundation model against matched classical features; inner speech in Greek and English, near its chance level; flat channels in a handwriting dataset, where masked EEG sits just above chance; and a stored-clip check on an emotion dataset, where simple controls have smaller errors than EEG. · Orthosis, 8 people · inner speech, BINGO, 17 subject fields · handwriting, 5 people · REFED, 2 nominal folders, 30 clips ### Which kinds of transfer have been measured What changes between training and test, and whether this site has evidence on it. | What changes | Measured | Status only | Held | Not measured | | --- | --- | --- | --- | --- | | Person | [Core matrix · new-person protocols](https://bci.report/#overview); [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) n=102 also changes: sensor; [LaBraM adaptation · EEGMAT](https://bci.report/topics/model-adaptation/#adaptation) n=36; [Standing to walking and running, SSVEP](https://bci.report/topics/on-the-move/) n=23 also changes: movement; [Pretraining controls](https://bci.report/topics/does-pretraining-help/) n=36 / 10; [Nine encoders adapted · EEGMAT](https://bci.report/topics/does-pretraining-help/#v9-adaptation) n=36; [Parkinson’s disease vs. controls](https://bci.report/topics/clinical-groups/) n=149; [Infantile spasms vs. controls · IESSEEG](https://bci.report/topics/clinical-groups/#iesseeg) n=100; [Schizophrenia vs. controls · ASEEG](https://bci.report/topics/clinical-groups/#aseeg) n=101; [Sleep staging · Dreem, two cohorts](https://bci.report/topics/sleep-staging/) n=55 / 25; [fNIRS, executed tapping · Bak et al.](https://bci.report/topics/beyond-scalp-eeg/#bak-fnirs) n=30; [Cursor control, dry electrodes · Shin2022](https://bci.report/topics/data-quality/#shin2022) n=10; [Orthosis device errors · EEG pilot](https://bci.report/topics/pilot-studies/#orthosis) n=8; [Inner speech, Greek and English · BINGO](https://bci.report/topics/pilot-studies/#bingo); [Imagined handwriting, four coded labels · EEG pilot](https://bci.report/topics/pilot-studies/#handwriting) n=5; [Frozen NeuroLM, arithmetic and rest · EEGMAT](https://bci.report/topics/pilot-studies/#neurolm-eegmat) n=36; [Frozen NeuroRVQ on three tasks](https://bci.report/topics/pilot-studies/#neurorvq) n=36 / 21 / 10 | | [Neural and foundation models on the Dreem cohorts](https://bci.report/releases/#hold-dreem-amplitude-sensitive-models); [Foundation models on the clinical cohort](https://bci.report/releases/#hold-clinical-foundation-models) | | | Session / day | [Mobile ERP · first session to the others](https://bci.report/topics/on-the-move/#erp-heading) n=24 / 17 also changes: movement; [OpenBMI motor imagery · first session to the second](https://bci.report/topics/calibration-budget/#next-session) n=51; [WBCIC-SHU motor imagery · an earlier session to a later one](https://bci.report/topics/later-sessions/#wbcic-shu) n=51 / 11; [Longitudinal RSVP · first visit to later visits](https://bci.report/topics/later-sessions/#longitudinal-rsvp) n=15; [Continuous cursor tracking · earliest to latest session](https://bci.report/topics/later-sessions/#forenzo) | [Cross-session pilot](https://bci.report/topics/model-adaptation/#cross-session) n=1 | [Next-day adaptation](https://bci.report/releases/#hold-bnci2015-001-crossday) | | | Sensor / montage | [Dry ⇄ wet electrodes](https://bci.report/topics/dry-vs-wet/) n=102 also changes: person; [In-ear vs. scalp, sleep](https://bci.report/topics/fewer-electrodes/#montage-finding) n=10 comparison; [Posterior subset vs. all electrodes](https://bci.report/topics/fewer-electrodes/#posterior-subset) n=19 comparison; [BETA, eight vs. four electrodes](https://bci.report/topics/fewer-electrodes/#v9-montage) n=70 comparison; [Cap shifted by one electrode spacing, BETA](https://bci.report/topics/fewer-electrodes/#cap-shift) n=70 | | | From one headset or amplifier to another | | Context | [Screen ⇄ VR, P300](https://bci.report/topics/screen-to-vr/) n=21; [Standing to walking and running, SSVEP](https://bci.report/topics/on-the-move/) n=23 also changes: person; [Standing to walking and running, ERP](https://bci.report/topics/on-the-move/#erp-heading) n=24 / 17 also changes: session; [Image rate, P300](https://bci.report/topics/screen-to-vr/#image-rate) n=9 also changes: recording; [Greek ⇄ English inner speech](https://bci.report/topics/pilot-studies/#bingo) also changes: person; [SSVEP frequencies left out of training · BETA](https://bci.report/topics/new-ssvep-frequencies/) n=70 also changes: person | | | | | Dataset | [BETA to Wearable-102, SSVEP](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer) n=102 also changes: sensor | | | | n is the number of people behind an entry; two sizes (n=24 / 17) are two analyses or cohorts in one entry, largest first. “comparison”: two electrode sets scored separately on the same data. “also changes”: what else changed in the same contrast. [The full map, with what each entry measures →](https://bci.report/topics/#transfer) Holds ## What is held back, and why - [Kim beta-range SSVEP · amplitude-dependent models](https://bci.report/topics/data-quality/#kim2025-beta-range): Held · Amplitude units, channel gains and the saved reference history are not established, so models whose output depends on amplitude are held on these files. Training-free CCA, which does not depend on a common amplitude scale, is published. - [Damm finger motor imagery · one-person score](https://bci.report/topics/data-quality/#damm): Status only · One participant’s two random-order runs, checked end to end. The scores are held under the one-person rule; design, counts and file facts are published. - [NeBULA · one-person scores](https://bci.report/topics/data-quality/#nebula): Status only · One participant reaching to three targets under three levels of assistance. The EEG, EMG and combined scores are held under the one-person rule; design, counts and file facts are published. - [ZuCo on NEMAR · converted-run timing](https://bci.report/topics/data-quality/#zuco2): No score · In one converted run on NEMAR, some fixation times fall outside the recording. No EEG–text score is computed until the converter’s timing rules and a source-checksummed fixture are available. - [ds007420 · fNIRS classifier](https://bci.report/topics/beyond-scalp-eeg/#ds007420): No score · Its numeric task codes have no verified lookup to hands or movements, and its event durations disagree between the events table, the optical files and the cited paper. Its optical coverage counts are published. - [Dreem sleep cohorts · neural-network and foundation models](https://bci.report/topics/sleep-staging/#held): Held · The stored signal metadata says millivolts; the publisher’s own converter treats the same arrays as microvolts. Models that need absolute amplitude are not run; the published baselines do not depend on the unit. - [Stieger longitudinal BCI · one-person pilot](https://bci.report/topics/model-adaptation/#cross-session): Status only · The pipeline works on one person’s eleven sessions. With one person every score is that person’s, so none is published. - [OpenNeuro ds004902 · paired sleep comparison](https://bci.report/releases/#hold-ds004902): No score · Five source payloads are shorter than their own headers declare. - [Foundation models on the clinical set](https://bci.report/topics/clinical-groups/): Held · Not run: the source states no physical amplitude unit. - [L-FAME](https://bci.report/releases/#hold-lfame): Described, not scored · Its per-condition numbers are withheld. - [BNCI2015-001 · next-day adaptation](https://bci.report/topics/model-adaptation/#next-day): Held · Run and independently replayed. The catalogue licence is CC BY-NC-ND and the description names no ethics approval, so no figure is published until that review is done. [Full holds register →](https://bci.report/releases/#holds) [Corrections (4) →](https://bci.report/releases/#corrections) Directory ## Everything else, by what it is - [22 methods with result pages](https://bci.report/methods/): Methods · Every published figure for each method; the model directory, with why a catalogued model has no score; external field notes. - [32 datasets with results](https://bci.report/datasets/): Datasets · Every figure measured on each dataset; the public-data register, with source terms, credit and what no release includes. - [36 reviewed releases](https://bci.report/releases/): Releases · What the newest releases changed, a line for every earlier one, corrections, the holds register, and an archive with every file’s SHA-256. - [72 downloadable files](https://bci.report/api/): Data & API · Every JSON and CSV file with what it holds, code examples, the release feed and how to cite. Evaluation ## How these results were produced Only explicitly reviewed cohort-level research results are exported. Raw EEG, individual results and model weights stay outside the website. Each comparison states its source, license, protocol and limitations. No cross-task overall score. [Download all results · JSON ↓](https://bci.report/data/experiments.json) 01 ### Status labels Downloads, loading checks and scored evaluations have distinct status labels. Models awaiting an adapter have no result. 02 ### Failure modes Missed commands, false activations and abstention are reported. 03 ### Same task, same protocol Splits, channels, preprocessing and training modes travel with each protocol. Frozen encoders and scratch baselines are labeled separately. 04 ### Fixed, audited local runs Experiments run offline on a Mac or local GPU workstation. Timings are configuration-specific. This website performs no EEG inference or diagnosis. --- Markdown copy of https://bci.report/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # What the evidence can answer Questions Answers from published decoding results, each with its cohort, interval and limits. Figures under one question often share participants: they are repeated conditions within its protocols. ## Transfer Does a decoder still work after a change of sensor, display, electrode layout, body movement or session? - [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/): Sensor transfer · What changes when a decoder crosses between two native eight-channel recordings from the same 102 people? · 2 s SSVEP · 12 targets · balanced accuracy - [Screen to VR](https://bci.report/topics/screen-to-vr/): Context transfer · The same 21 people calibrated on a PC screen and tested in a VR headset, and the reverse: the level reached after the switch, with no same-display score to compare. Plus: a P300 decoder trained at one image rate and tested at another. · P300 · 21 people in VR · 9 people across image rates - [Fewer electrodes](https://bci.report/topics/fewer-electrodes/): Montage · In-ear against scalp for sleep, four posterior electrodes against sixteen for eyes open or closed, foundation encoders on BETA with eight and four electrodes, and a simulated cap shift of one electrode spacing on BETA: channel sets from specific research recordings. · Two paired comparisons · 10 and 19 people · BETA, 8 and 4 electrodes, 70 people · a simulated cap shift - [On the move](https://bci.report/topics/on-the-move/): Motion robustness · Standing, walking and running results, with scalp and ear recordings and incompatible time windows kept apart. · SSVEP balanced accuracy · ERP ROC AUC - [Later sessions](https://bci.report/topics/later-sessions/): Session transfer · same person · Trained on a person’s earlier session and tested on a later one, with no labels from it: motor imagery across recording sessions, with simple baselines and frozen CBraMod; target detection at the publisher’s nominal later visits; and continuous cursor tracking, where a constant baseline does better. · Motor imagery, WBCIC-SHU, 51 and 11 people · RSVP, 15 people · cursor tracking, a negative result ## Adapting models How much calibration a decoder needs, which part of a pretrained model to update, whether pretraining helps, whether one model can answer several questions, including questions asked in words, and what lets a decoder answer an SSVEP frequency it never trained on. - [How much calibration?](https://bci.report/topics/calibration-budget/): Calibration budget · Twelve, 24 or 48 labeled trials on a wearable SSVEP task, and up to 40 from a person’s next motor-imagery session: the gain differs by method, and some people’s scores fall. · SSVEP · 102 people · motor imagery, next session · 51 people - [Which part to update?](https://bci.report/topics/model-adaptation/): Model adaptation · LaBraM on new people, same task, zero labels from the test person: head only, last block or LoRA, printed beside the core matrix’s frozen readout. Plus: the next-day experiment, run and held. · EEGMAT, 36 people · three update rules · next day: held - [Does pretraining help?](https://bci.report/topics/does-pretraining-help/): Representation controls · Matched pretrained and constructor-random encoders under fixed and train-selected readout settings, and the same control for five further checkpoints on three tasks. Plus further foundation models in the same frozen recipe: three checkpoints and, narrowly, EEGPT above every published row on sleep staging, none above training-free CCA on BETA. · Three tasks · eight encoders · three random initializations · 23 further checkpoints - [One model, several questions](https://bci.report/topics/shared-encoder/): Shared encoder · route 2 · One shared encoder with a fixed output per question, against a head told which question it is answering and against a separate model per question, at matched data and compute — with what each costs. · Motor imagery, OpenBMI, 51 people · sleep, BOAS · EESM19, crude - [Questions in language](https://bci.report/topics/questions-in-language/): Questions in language · route 3 · One small head over frozen EEG features, asked each question by a number, a label template or a description: it works only for questions it was trained on, and not on every dataset. · SSVEP, BETA, 70 people · sleep, BOAS · secondary: EESM19, OpenBMI, Wearable-102 - [New SSVEP frequencies](https://bci.report/topics/new-ssvep-frequencies/): Route 3 follow-ups · Fourteen follow-ups: thirteen on route 3’s BETA cohort, five on Wearable-102, one of them only there. Heads that read reference spectra or stimulus phase stayed far below CCA, which needs no training; a small learned correction on top of CCA added a few points on BETA and changed little on Wearable-102; filter-bank CCA, untrained too, scored above CCA in every setting tested, and an untrained harmonic sum came out nearly level with a learned eigenvalue head. · SSVEP, BETA, 70 people · 8 held-out frequencies · Wearable-102, 102 people ## Reliability & clinical When a decoder should not act, what accuracy hides in sleep staging, what resting-state EEG says about a clinical group, and what a public dataset’s files show when checked end to end. - [When not to act](https://bci.report/topics/when-not-to-act/): Abstention · decision research · Command detection read with its false-activation rate; when a decoder should decline to decide, on two protocols; and the decision-research plan, with its three routes. · Idle, 4-person pilot · non-control, 20 people · reliable decisions, 36 and 70 people - [Sleep-stage balance](https://bci.report/topics/sleep-staging/): Sleep staging · simple baselines · Two simple baselines in 25 healthy sleepers and 55 people with sleep apnoea, as two separate experiments: accuracy beside balanced accuracy, and the stage the stager never predicts. · Dreem DOD-H and DOD-O · 25 and 55 nights · balanced accuracy - [Clinical groups](https://bci.report/topics/clinical-groups/): Clinical research · A 149-person Parkinson's and control comparison beside an age-and-sex-only comparator; neither is a diagnosis. · 149 people · one site · balanced accuracy - [Checked end to end](https://bci.report/topics/data-quality/): Data quality and reproducibility · Public EEG datasets read from their own files: Shin2022’s runs, duplicates and valid windows counted before a participant-held-out check, where the EEG arm sits close to a control that reads only trial duration; training-free CCA on Kim beta-range SSVEP and on every eldBETA participant, with filter-bank CCA beside it there; MIND’s stored markers, read on record IDs; two one-person cases, Damm and NeBULA, with counts and no score; and a timing hold on one converted ZuCo run. · Shin2022, 10 people · Kim beta-range · MIND · eldBETA, 100 people · one-person cases, counts only · ZuCo, status only ## Other signals & pilots What simple baselines show on signals other than scalp EEG, in people and in one macaque, and what small pilot studies show. - [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/): Other signals · human and macaque · fNIRS telling right-hand, left-hand and foot tapping apart in people the classifier never saw; finger velocity from one macaque’s motor-cortex spiking, where a ridge decoder lowers the squared error but not the absolute error; and a numerical coverage check of fNIRS recordings. Intracortical handwriting and functional ultrasound: status only. · fNIRS, 30 people · intracortical spiking, 1 macaque · fNIRS coverage, 12 people - [Pilot studies](https://bci.report/topics/pilot-studies/): Pilot studies · small cohorts · Small pilots: injected orthosis errors in one cohort, including a frozen EEG foundation model against matched classical features; inner speech in Greek and English, near its chance level; flat channels in a handwriting dataset, where masked EEG sits just above chance; and a stored-clip check on an emotion dataset, where simple controls have smaller errors than EEG. · Orthosis, 8 people · inner speech, BINGO, 17 subject fields · handwriting, 5 people · REFED, 2 nominal folders, 30 clips [Jev-style questions on EEG: the evidence](https://bci.report/jev-style/): Decision research · “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. Three routes, each answered by this site’s own runs on public EEG data. · Read the evidence → Transfer coverage ## Which kinds of transfer have been measured What changes between training and test, and whether this site has evidence on it. | What changes | Measured | Status only | Held | Not measured | | --- | --- | --- | --- | --- | | Person — New people: trained on others, tested on someone the model never saw | [Core matrix · new-person protocols](https://bci.report/#overview) — Every protocol whose split is “transfer to a new person”: participant-disjoint folds.; [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) n=102 also changes: sensor — Source models trained on other people, so the new person and the new sensor arrive together.; [LaBraM adaptation · EEGMAT](https://bci.report/topics/model-adaptation/#adaptation) n=36 — New people, same task, zero labels from the test person.; [Standing to walking and running, SSVEP](https://bci.report/topics/on-the-move/) n=23 also changes: movement — Trained on other people standing, tested on a new person on the move.; [Pretraining controls](https://bci.report/topics/does-pretraining-help/) n=36 / 10 — Frozen encoders with a readout fitted on other people, on a motor-imagery and a mental-workload task.; [Nine encoders adapted · EEGMAT](https://bci.report/topics/does-pretraining-help/#v9-adaptation) n=36 — A head on the frozen encoder against LoRA, new people, zero labels from the test person, one fixed recipe.; [Parkinson’s disease vs. controls](https://bci.report/topics/clinical-groups/) n=149 — Every person held out once; one site.; [Infantile spasms vs. controls · IESSEEG](https://bci.report/topics/clinical-groups/#iesseeg) n=100 — The publisher’s participant-disjoint folds: each person held out once; one institution.; [Schizophrenia vs. controls · ASEEG](https://bci.report/topics/clinical-groups/#aseeg) n=101 — Each person held out once in stratified participant folds, averaged over split seeds; one clinic.; [Sleep staging · Dreem, two cohorts](https://bci.report/topics/sleep-staging/) n=55 / 25 — Each night held out once and scored by a model fitted on other people’s nights of the same cohort; the two cohorts are separate experiments.; [fNIRS, executed tapping · Bak et al.](https://bci.report/topics/beyond-scalp-eeg/#bak-fnirs) n=30 — Human fNIRS: each person held out once and classified by a model fitted on everyone else.; [Cursor control, dry electrodes · Shin2022](https://bci.report/topics/data-quality/#shin2022) n=10 — Each participant held out once and scored by a ridge fitted on the others.; [Orthosis device errors · EEG pilot](https://bci.report/topics/pilot-studies/#orthosis) n=8 — Each person held out once, with a probe fitted on the other seven: classical features on two channel sets and frozen EEGPT, one cohort counted once.; [Inner speech, Greek and English · BINGO](https://bci.report/topics/pilot-studies/#bingo) — A negative result near chance, a subject field left out at a time. Counted in subject fields as the source defines them, not proven distinct people, so no n is shown.; [Imagined handwriting, four coded labels · EEG pilot](https://bci.report/topics/pilot-studies/#handwriting) n=5 — Each person held out once, with a model fitted on the other four: a reused development cohort and a data-quality case study, just above chance.; [Frozen NeuroLM, arithmetic and rest · EEGMAT](https://bci.report/topics/pilot-studies/#neurolm-eegmat) n=36 — Whole participants held out across folds; a separate pilot panel.; [Frozen NeuroRVQ on three tasks](https://bci.report/topics/pilot-studies/#neurorvq) n=36 / 21 / 10 — Whole participants held out on each task; a separate pilot panel. | | [Neural and foundation models on the Dreem cohorts](https://bci.report/releases/#hold-dreem-amplitude-sensitive-models) — Not run: the source’s physical units disagree.; [Foundation models on the clinical cohort](https://bci.report/releases/#hold-clinical-foundation-models) — Not run: the source states no physical amplitude unit. | | | Session / day — The same person, another session or another day | [Mobile ERP · first session to the others](https://bci.report/topics/on-the-move/#erp-heading) n=24 / 17 also changes: movement — Fitted on the person’s first session, standing; the session and the movement change together.; [OpenBMI motor imagery · first session to the second](https://bci.report/topics/calibration-budget/#next-session) n=51 — The same person and task: trained on the first session, tested on the second, with and without a few of its labelled trials.; [WBCIC-SHU motor imagery · an earlier session to a later one](https://bci.report/topics/later-sessions/#wbcic-shu) n=51 / 11 — Two cohorts kept apart: simple baselines, and frozen CBraMod with a readout, fitted on each person’s first recording session and tested on their third with no labels from it.; [Longitudinal RSVP · first visit to later visits](https://bci.report/topics/later-sessions/#longitudinal-rsvp) n=15 — One decoder per person, trained at the first visit and scored at the publisher’s nominal later visits.; [Continuous cursor tracking · earliest to latest session](https://bci.report/topics/later-sessions/#forenzo) — A negative result: a constant baseline did better. Counted in records, not proven unique people, so no n is shown. | [Cross-session pilot](https://bci.report/topics/model-adaptation/#cross-session) n=1 — One person: every score would be that person’s, so none is published. | [Next-day adaptation](https://bci.report/releases/#hold-bnci2015-001-crossday) — Run and independently replayed; held for a licence and ethics review. | | | Sensor / montage — Another sensor type or another set of electrodes | [Dry ⇄ wet electrodes](https://bci.report/topics/dry-vs-wet/) n=102 also changes: person — Trained on one sensor type, tested on the other, in both directions.; [In-ear vs. scalp, sleep](https://bci.report/topics/fewer-electrodes/#montage-finding) n=10 comparison — Both electrode sets scored on the same epochs.; [Posterior subset vs. all electrodes](https://bci.report/topics/fewer-electrodes/#posterior-subset) n=19 comparison — A software subset of one headset.; [BETA, eight vs. four electrodes](https://bci.report/topics/fewer-electrodes/#v9-montage) n=70 comparison — Two core protocols on the same people and folds, each trained and tested on its own.; [Cap shifted by one electrode spacing, BETA](https://bci.report/topics/fewer-electrodes/#cap-shift) n=70 — Simulated on the same recordings: decoders fitted at the original positions, tested with each electrode read one position over. | | | From one headset or amplifier to another — Every sensor change here is within one study’s own equipment. | | Context — Another display, movement, image rate, language or stimulus frequency | [Screen ⇄ VR, P300](https://bci.report/topics/screen-to-vr/) n=21 — The same people, calibrated on one display and tested on the other.; [Standing to walking and running, SSVEP](https://bci.report/topics/on-the-move/) n=23 also changes: person — Trained on other people standing, tested on a new person on the move.; [Standing to walking and running, ERP](https://bci.report/topics/on-the-move/#erp-heading) n=24 / 17 also changes: session — The same person; movement and session change together.; [Image rate, P300](https://bci.report/topics/screen-to-vr/#image-rate) n=9 also changes: recording — Trained on one recording at one rate, tested on a different recording: rate and recording change together.; [Greek ⇄ English inner speech](https://bci.report/topics/pilot-studies/#bingo) also changes: person — Trained in one language and tested in the other, on a subject field left out: the language and the person change together. A negative result near chance.; [SSVEP frequencies left out of training · BETA](https://bci.report/topics/new-ssvep-frequencies/) n=70 also changes: person — Tested on frequencies and people left out of training together. | | | | | Dataset — Trained on one dataset, tested on another | [BETA to Wearable-102, SSVEP](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer) n=102 also changes: sensor — Models trained on BETA, run unchanged on Wearable-102’s dry and wet headbands; the documentation cannot rule out that some people are in both. | | | | n is the number of people behind an entry; two sizes (n=24 / 17) are two analyses or cohorts in one entry, largest first. “comparison”: two electrode sets scored separately on the same data. “also changes”: what else changed in the same contrast. Cohort sizes are read from the reviewed downloads: [deployment-topics.json](https://bci.report/data/deployment-topics.json) · [adaptation-update.json](https://bci.report/data/adaptation-update.json) · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) · [clinical-update.json](https://bci.report/data/clinical-update.json) · [clinical-iesseeg-update.json](https://bci.report/data/clinical-iesseeg-update.json) · [clinical-aseeg-update.json](https://bci.report/data/clinical-aseeg-update.json) · [large-source-update.json](https://bci.report/data/large-source-update.json) · [beyond-scalp-update.json](https://bci.report/data/beyond-scalp-update.json) · [data-quality-update.json](https://bci.report/data/data-quality-update.json) · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) · [neurolm-eegmat-update.json](https://bci.report/data/neurolm-eegmat-update.json) · [neurorvq-update.json](https://bci.report/data/neurorvq-update.json) · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) · [context-update.json](https://bci.report/data/context-update.json) · [evidence-update.json](https://bci.report/data/evidence-update.json) · [electrode-shift-update.json](https://bci.report/data/electrode-shift-update.json) · [extension-update.json](https://bci.report/data/extension-update.json) · [new-ssvep-frequencies-update.json](https://bci.report/data/new-ssvep-frequencies-update.json) · [Full holds register →](https://bci.report/releases/#holds) --- Markdown copy of https://bci.report/topics/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Do dry EEG electrodes decode as well as wet ones? Sensor transfer · 102 participants Balanced accuracy falls most when an EEGNet trained on wet recordings is tested on dry recordings from held-out people. - [Reviewed aggregate JSON ↓](https://bci.report/data/deployment-topics.json) - Measured on: [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Methods: [EEGNet](https://bci.report/methods/eegnet/); [CCA](https://bci.report/methods/cca/); [FBCCA](https://bci.report/methods/fbcca/) ## Short answer Not in this dataset. On a 12-target, two-second SSVEP task, with the same **102** people recorded on both sensor types, the training-free FBCCA reference scored **82.3%** on wet recordings and **64.6%** on dry ones. An EEGNet trained on wet recordings fell by **29.4 pp** of balanced accuracy when tested on dry recordings (paired 95% interval **−34.5 pp** to **−24.3 pp**). Wearing order and time on task change together with sensor type, so the hardware’s own share of the gap cannot be separated. **Wet and dry recordings of the same 102 people.** Participant-mean balanced accuracy on 12-target, two-second SSVEP. Dot: the estimate; line: participant bootstrap 95% interval; dashed line: chance, 1 in 12. Dark rows are wet recordings, rust rows dry. - FBCCA · wet recording — training-free: 82.3% (78.7%–85.7%) - FBCCA · dry recording — training-free: 64.6% (59.5%–69.6%) - EEGNet · wet → wet — trained on the first sensor, tested on the second: 75.0% (70.3%–79.5%) - EEGNet · wet → dry — trained on the first sensor, tested on the second: 45.6% (40.5%–50.8%) - EEGNet · dry → dry — trained on the first sensor, tested on the second: 56.8% (51.9%–61.6%) - EEGNet · dry → wet — trained on the first sensor, tested on the second: 64.0% (59.1%–68.8%) **−29.4 pp** Wet-trained EEGNet: dry minus wet balanced accuracy ### The largest observed cross-sensor loss Across the same 102 people, the paired participant bootstrap 95% interval is −34.5 to −24.3 pp. Source-trained models ## Train on one sensor, test on both Six folds hold out 17 people at a time. Both sensor recordings from each held-out person stay out of source training. Every cell is participant-mean balanced accuracy on 12-target, two-second SSVEP. Native 8-channel headband · 102 participants · 12,240 trials per target sensor · participant bootstrap 95% intervals. | Model and training source | Test on wet | Test on dry | | --- | --- | --- | | Spectral ridge — trained on wet · participant-disjoint | 40.3% (36.5%–44.3%) | 35.2% (31.8%–38.8%) | | Spectral ridge — trained on dry · participant-disjoint | 39.7% (35.9%–43.6%) | 37.2% (33.5%–41.0%) | | EEGNet — trained on wet · participant-disjoint | 75.0% (70.3%–79.5%) | 45.6% (40.5%–50.8%) | | EEGNet — trained on dry · participant-disjoint | 64.0% (59.1%–68.8%) | 56.8% (51.9%–61.6%) | Wet-trained EEGNet 75.0% → 45.6% Dry-trained EEGNet 56.8% → 64.0% Testing the dry-trained model on wet recordings increased balanced accuracy by 7.3 points; 95% interval +4.2 to +10.4 pp. A different kind of reference 82.3% / 64.6% FBCCA reads each target recording analytically, with no training. Analytic references ## CCA, FBCCA and a power reference Training-free: each scores the target recording directly. Target-recording references on the same two-second, 12-target protocol. | Reference | Wet recording | Dry recording | | --- | --- | --- | | Same-frequency power — analytic or fixed reference | 39.8% · 95% interval 35.4%–44.3% | 36.0% · 95% interval 31.8%–40.2% | | CCA — analytic or fixed reference | 81.6% · 95% interval 77.7%–85.1% | 65.5% · 95% interval 60.6%–70.2% | | FBCCA — analytic or fixed reference | 82.3% · 95% interval 78.7%–85.7% | 64.6% · 95% interval 59.5%–69.6% | Methods & limits ## Limits ### Order remains a confound Wearing order and time on task change with sensor type. Both order strata show the wet-trained drop. ### Amplitude units are unresolved The release documents no physical amplitude unit, so the scale-dependent foundation-model runs remain held. ### One EEGNet seed The source-trained EEGNet cells use one training seed. The paired intervals describe participant variation conditional on that fitted protocol. ### An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces. Sensors, 21(4), 1256. https://doi.org/10.3390/s21041256 Aggregate results computed from the official Tsinghua BCI Lab author mirror snapshot acquired 2026-09-20. [Dataset record ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) · [Sensors paper ↗](https://doi.org/10.3390/s21041256) · CC BY 4.0 Data source: [reviewed aggregate JSON](https://bci.report/data/deployment-topics.json) · schema bci-report-public-deployment-topics-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Do dry EEG electrodes decode as well as wet ones?* https://bci.report/topics/dry-vs-wet/ Figures from release [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/dry-vs-wet/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Does a P300 decoder calibrated on a screen still work in VR? Context transfer · P300 on a screen and in VR 21 people used the same P300 interface on a PC screen and in a virtual-reality headset, calibrated on one display and tested on the other. - [Reviewed aggregate JSON ↓](https://bci.report/data/context-update.json) - Measured on: [Cattan PC/VR P300 dataset](https://bci.report/datasets/vr-pc-p300/); [LTRSVP · EEG Signals from an RSVP Task](https://bci.report/datasets/ltrsvp/) ## Short answer It scores above chance after the switch. Across **21** people, each calibrated on a PC screen and tested in a VR headset and also the other way round, a simple logistic regression averaged **66.0%** balanced accuracy over the two directions (95% interval **61.6%–70.2%**; chance **50%**), with stimulus-onset timing corrected. The protocol did not also test each person on the display they were calibrated on, so the cost of the switch is unmeasured. **The level after a display change, 21 people.** Each person was calibrated on one display and tested on the other, in both directions, averaged within the person. Balanced accuracy, onset-corrected timing. Dot: mean over people; line: participant-bootstrap 95% interval; dashed line: chance, 50%. - Mean-window logistic regression: 66.0% (61.6%–70.2%) - Spatiotemporal shrinkage LDA: 63.6% (59.6%–67.7%) **66.0%** Balanced accuracy after the display change, 21 people ### Above chance after the display change Interval 61.6%–70.2%; AUROC 0.732. Two fixed baselines ## Calibrated on one display, tested on the other Target against non-target epochs, 5 non-targets to every target, so accuracy would reward always answering “non-target”; balanced accuracy and AUROC do not. Primary timing. Mean over 21 people of each person’s average of the two transfer directions, with participant-bootstrap 95% intervals. | Baseline | Balanced accuracy | AUROC | | --- | --- | --- | | Mean-window logistic regression | 66.0% (61.6%–70.2%) | 0.732 (0.671–0.785) | | Spatiotemporal shrinkage LDA | 63.6% (59.6%–67.7%) | 0.713 (0.652–0.771) | Prespecified sensitivity ## With and without the onset correction The headset and the screen report stimulus onset with different delays. The primary analysis shifts each epoch by the source-reported offset — 19 samples on the PC and 60 in VR, at 512 Hz. The sensitivity analysis leaves epochs at the recorded tag. Balanced accuracy under each timing scheme, with participant-bootstrap 95% intervals. | Baseline | Onset corrected (primary) | At the recorded tag | | --- | --- | --- | | Mean-window logistic regression | 66.0% (61.6%–70.2%) | 63.8% (61.0%–66.5%) | | Spatiotemporal shrinkage LDA | 63.6% (59.6%–67.7%) | 57.2% (55.1%–59.4%) | Spatiotemporal shrinkage LDA depends on timing most: 63.6% with the correction, 57.2% without. What the scores are computed on ## Twenty-one people, one recording on each display Every label comes from the source. One malformed row after the last epoch of one recording was discarded. People 21 each recorded once on each display Labelled events 30,240 1 target to 5 non-targets Epochs kept, primary timing 30,101 139 rejected at 500 µV peak to peak Another presentation change · image rate ## Trained at 5 images a second, tested at 10: does a P300 decoder carry over? In LTRSVP, a separate dataset scored with a different decoder, 9 people watched aerial images flash past at 5, 6 and 10 a second and looked for the ones with an airplane, recording twice at each rate. A fixed logistic regression is trained on one recording at one rate (run a) and tested on a different recording of the same person (run b), at the same rate or another; within a rate, run b followed run a after a long break. Both arms below are tested on the same 10-Hz images: 223 targets and 2,036 non-targets across the 9 people. Trained at 5 Hz, run a → tested at 10 Hz, run b 60.8% AUROC 0.673 (0.617–0.732) — 95% interval 56.5%–65.8% Trained at 10 Hz, run a → tested at 10 Hz, run b 63.0% AUROC 0.695 (0.630–0.753) — 95% interval 57.7%–68.0% **−2.2 pp** Trained at 5 Hz minus trained at 10 Hz, same 10-Hz test images, 9 people ### Slightly lower on average — and the interval crosses zero Paired interval −6.8 to +2.8 pp: it crosses zero, so no change is established. 7 of 9 people had a lower point estimate and 2 a higher one. AUROC moved by −0.022 (interval −0.060 to +0.014), which also crosses zero. Participant-mean balanced accuracy for each training rate (rows, run a) and test rate (columns, run b), 9 people, with participant-bootstrap 95% intervals. Chance is 50%. | Training recording | Tested at 5 Hz — 338 targets / 3,069 non-targets | Tested at 6 Hz — 296 targets / 2,664 non-targets | Tested at 10 Hz — 223 targets / 2,036 non-targets | | --- | --- | --- | --- | | Trained at 5 Hz | 73.6% (65.8%–80.8%) | 70.7% (64.5%–76.7%) | 60.8% (56.5%–65.8%) | | Trained at 6 Hz | 74.9% (68.5%–80.9%) | 71.4% (64.3%–78.1%) | 60.6% (57.4%–64.3%) | | Trained at 10 Hz | 67.5% (59.9%–75.2%) | 69.8% (61.9%–77.2%) | 63.0% (57.7%–68.0%) | Every cell in a column is tested on the same images, whatever the training rate. On the diagonal the test recording is the same-rate run b, after a long break; off it, the rate also changes, and so does how far apart in the session the two recordings were. The nine cells reuse the same people and the same three training recordings of each. ### Limits of the image-rate comparison ### Rate and recording are confounded Training at another rate also means training on another recording, with another number of images and another target balance. A difference here is not a causal effect of image rate. ### Ascending presentation order The original publication reports that the rates were presented from the lowest to the highest, in an order not randomised across participants. Within a rate, run b is the later recording, after a long break; how the two files of each rate map onto the ascending sequence is not documented on PhysioNet. So a cross-rate cell also differs in elapsed time, fatigue and practice. ### 9 people The 8,626 test images and 81 person-by-cell evaluations all come from these 9 people, and the intervals are bootstrapped over people. 2 of the release's 11 people lack both recordings at one rate and were excluded from file metadata before scoring. ### Overlapping responses, offline At these rates one image’s baseline and response windows also contain its neighbours’ responses. All results are offline, from a fixed decoder. ### LTRSVP · EEG Signals from an RSVP Task Ana Matran (University of Essex) · EEG Signals from an RSVP Task 1.0.0, PhysioNet (2017), doi:10.13026/C2KX0P. Original publication: Ana Matran-Fernandez and Riccardo Poli, Towards the automated localisation of targets in rapid image-sifting by collaborative brain-computer interfaces, PLoS ONE 12(5): e0178498 (2017), doi:10.1371/journal.pone.0178498. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. The original publication states that the study was approved by the Ethics Committee of the University of Essex and that every participant gave written consent before the experiment. Only cohort aggregates appear here, not the ages, sexes or handedness the paper reports. [PhysioNet record ↗](https://physionet.org/content/ltrsvp/1.0.0/) · [Original publication ↗](https://doi.org/10.1371/journal.pone.0178498) · [Open Data Commons Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/) Data source: [extension-update.json](https://bci.report/data/extension-update.json) · schema bci-report-extension-update-v1. Methods & limits ## Limits of the display-transfer result ### No same-display reference Nobody was tested on the display they were calibrated on. ### Within one person, on one day Each person supplies both their calibration and their test data, on the same day. ### Simple baselines only Two simple CPU classifiers with no tuning; no foundation model or fine-tuned model was run. ### Event-to-event jitter The correction shifts epochs by the mean onset delay; jitter around that mean, which a headset can add, remains. ### Cattan PC/VR P300 · a BCI experiment in virtual reality and on a personal computer Grégoire Cattan, Anton Andreev, Pedro L. C. Rodrigues and Marco Congedo · Dataset of an EEG-based BCI experiment in Virtual Reality and on a Personal Computer, Zenodo, doi:10.5281/zenodo.2605205; documentation arXiv:1903.11297. All participants gave written informed consent covering the experimental process, the data management procedures and the right to withdraw at any moment; the study was approved by the ethics committee of the University of Grenoble Alpes (CERNI). Only cohort aggregates appear here. [Zenodo record ↗](https://doi.org/10.5281/zenodo.2605205) · [Dataset documentation ↗](https://arxiv.org/abs/1903.11297) · [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/context-update.json) · schema bci-report-context-update-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Does a P300 decoder calibrated on a screen still work in VR?* https://bci.report/topics/screen-to-vr/ Figures from releases [`extension-update-20261002`](https://bci.report/releases/#extension-update-20261002) (2026-10-02), [`context-update-20260927`](https://bci.report/releases/#context-update-20260927) (2026-09-27). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/screen-to-vr/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Can fewer electrodes, or electrodes in the ear, match a full scalp montage? Montage · two paired comparisons and a simulated cap shift · BETA and sleep montages On the same 10 people and the same 30-second epochs, six scalp electrodes staged sleep better than four in-ear channels — for every one of the 10. A second comparison — four posterior electrodes against all sixteen for telling eyes open from closed — found no consistent difference across 19 people. On BETA, single-channel and channel-wise foundation encoders lost accuracy going from eight electrodes to four. A last section simulates the whole cap shifted by one electrode spacing on BETA. - [Reviewed aggregate JSON ↓](https://bci.report/data/evidence-update.json) - Measured on: [BETA](https://bci.report/datasets/beta/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [EESM23 · Ear-EEG Sleep Monitoring 2023](https://bci.report/datasets/eesm23/); [EEG Alpha Waves dataset](https://bci.report/datasets/alpha-waves/) · Methods: [EEGNet](https://bci.report/methods/eegnet/); [LaBraM](https://bci.report/methods/labram/); [CBraMod](https://bci.report/methods/cbramod/); [CCA](https://bci.report/methods/cca/); [REVE](https://bci.report/methods/reve/); [LUNA](https://bci.report/methods/luna/); [BrainOmni](https://bci.report/methods/brainomni/); [CodeBrain](https://bci.report/methods/codebrain/); [EEGMamba](https://bci.report/methods/eegmamba/); [ST-EEGFormer](https://bci.report/methods/st-eegformer/); [eeg-fm-masking](https://bci.report/methods/eeg-fm-masking/); [ERP-FM](https://bci.report/methods/erp-fm/); [SingLEM](https://bci.report/methods/singlem/); [ZUNA](https://bci.report/methods/zuna/); [EEGPT](https://bci.report/methods/eegpt/) ## Short answer In-ear channels did not, for sleep staging in this cohort: on the same **10** people and epochs, four in-ear channels scored **53.6%** balanced accuracy on five-stage sleep against **69.0%** for six scalp electrodes (chance **20%**), a paired gap of **+15.4 pp**. In a separate eyes-open-or-closed task on **19** other people, four posterior electrodes showed no consistent difference from all sixteen: all sixteen minus four came to **−1.6 pp** (paired 95% interval **−10.0 pp** to **+6.8 pp**). Both are channel sets from research recordings: the sleep gap mixes electrode position, count and hardware, and the four posterior electrodes are a software subset of one recording. On the core matrix's BETA protocol (**70** people), the further frozen foundation encoders show no small-montage advantage: the single-channel SingLEM went from **26.6%** with eight electrodes to **20.5%** with four and the channel-wise ZUNA 1.1 from **11.9%** to **8.4%**, and no new encoder lies above training-free CCA at four electrodes (**57.6%**). **Two separate paired comparisons, each on its own people and task.** Paired difference in person-mean balanced accuracy, scalp or all sixteen minus in-ear or four posterior. Dot: the mean difference; line: descriptive person-bootstrap 95% interval; dashed line: zero, no difference. Right of zero, the scalp or sixteen-electrode configuration scored higher. Chance is 20% for five-stage sleep and 50% for eyes open or closed. - Sleep staging · six scalp minus four in-ear · 10 people — 95% interval +10.6 to +20.3 pp: +15.4 pp - Eyes open or closed · all sixteen minus four posterior · 19 people — 95% interval −10.0 to +6.8 pp: −1.6 pp **+15.4 pp** Scalp minus in-ear balanced accuracy, same 10 people ### Every person scored higher on scalp 10 of 10 people scored higher with six scalp electrodes than with four in-ear channels. The descriptive person-bootstrap interval for the mean difference is +10.6 to +20.3 pp. Two configurations, identical epochs ## Four in-ear channels against six scalp electrodes Five-stage sleep, scored in 30-second epochs; chance is 20%. Ten leave-one-person-out folds and one fixed model for both configurations. Person-mean balanced accuracy and macro F1. Descriptive person-bootstrap 95% intervals below each balanced accuracy. | Configuration | Channels | Balanced accuracy | Macro F1 | | --- | --- | --- | --- | | Four in-ear channels — physical in-ear channels | RB · RT · LB · LT | 53.6% (47.4%–59.7%) | 47.6% | | Six scalp electrodes — mastoid slots M1 and M2 used only to screen for complete epochs | F3 · C3 · O1 · F4 · C4 · O2 | 69.0% (63.4%–74.2%) | 64.9% | Usable data ## What the scores are conditioned on An epoch was kept only if both configurations had complete, finite data for it; that rule dropped 3,849 epochs. The scores describe this subset. Paired epochs kept 75.2% 11,674 of 15,523 People 10 all retained Nights 17 17 of 19 Second comparison · eyes open and closed ## Four posterior electrodes against all sixteen A different task on different people: telling eyes open from eyes closed in 6-second windows, where chance is 50%. Both configurations are subsets of the same 16-electrode research recordings, scored with one fixed model over 19 leave-one-person-out folds. **−1.6 pp** All sixteen minus four posterior, same 19 people ### No consistent difference 7 of 19 people scored higher with all sixteen electrodes, 7 with the four posterior ones, and 5 tied. The descriptive person-bootstrap interval for the mean difference is −10.0 to +6.8 pp. It spans zero, which is not evidence that four electrodes are as good as sixteen. Person-mean balanced accuracy and macro F1. Descriptive person-bootstrap 95% intervals below each balanced accuracy. | Configuration | Channels | Balanced accuracy | Macro F1 | | --- | --- | --- | --- | | Four posterior electrodes — selected in software from the same recording | Pz · O1 · Oz · O2 | 79.5% (72.6%–85.8%) | 76.6% | | All sixteen electrodes — every electrode the cap recorded | Fp1 · Fp2 · Fc5 · Fz · Fc6 · T7 · Cz · T8 · P7 · P3 · Pz · P4 · P8 · O1 · Oz · O2 | 77.9% (70.5%–85.3%) | 74.5% | Third comparison · core-matrix protocols ## Eight or four electrodes on BETA, and six-channel sleep: the new frozen encoders The core matrix scores BETA’s 40-target SSVEP twice, with eight and with four posterior electrodes selected from the same 64-channel recordings (70 people, chance 2.5%), and sleep staging on six scalp channels (EESM19 scalp subset, 20 people, chance 20.0%). Eleven further foundation models joined these protocols as frozen probes, with the published recipe and only the encoder swapped, and so did EEGPT, in the same runs as the released LaBraM and CBraMod rows. Unlike the two comparisons above, each montage here is trained and tested on its own, and the released and the new rows share the people and folds. ### No small-montage advantage for single-channel or channel-wise encoders SingLEM, which encodes each channel on its own, went from 26.6% with eight electrodes to 20.5% with four (−6.1 pp); ZUNA 1.1, a channel-wise encoder, from 11.9% to 8.4% (−3.5 pp). For both, the two intervals do not overlap. SingLEM lost about as much as the montage encoders that lie above CBraMod at both montages (REVE Large −6.4 pp, eeg-fm-masking MAE (r=9 cm, L=2) −7.7 pp); ZUNA 1.1 started from a much lower score. On six-channel sleep SingLEM is the lowest new row (35.3%, 33.7%–37.0%), below every other row, released or new. ### No new encoder lies above standard CCA at four electrodes Standard CCA, which needs no training, scores 57.6% with four electrodes (51.8%–63.5%). No new row lies above it. Of the new rows, only eeg-fm-masking MAE (r=9 cm, L=2) has an interval that overlaps it (47.1%); of the released rows, so does spectral ridge (48.2%, 43.1%–53.3%). Frozen encoders with a ridge head beside the released rows · balanced accuracy with its descriptive 95% interval. | Method | BETA, 8 electrodes | BETA, 4 electrodes | Eight to four | Sleep, 6 channels | | --- | --- | --- | --- | --- | | Released rows · core matrix | | | | | | [Standard CCA](https://bci.report/methods/cca/) | 63.1% (57.2%–69.0%) | 57.6% (51.8%–63.5%) | Intervals overlap | — Not in the sleep protocol | | Spectral ridge | 50.7% (45.6%–55.8%) | 48.2% (43.1%–53.3%) | Intervals overlap | 72.2% (70.7%–73.8%) | | [LaBraM](https://bci.report/methods/labram/) | 10.8% (9.5%–12.2%) | 12.9% (11.2%–14.6%) | Intervals overlap | 71.1% (69.0%–73.2%) | | [CBraMod](https://bci.report/methods/cbramod/) | 33.7% (29.8%–37.6%) | 27.6% (24.1%–31.3%) | Intervals overlap | 71.3% (68.7%–73.7%) | | [EEGNet](https://bci.report/methods/eegnet/) | 55.8% (50.1%–61.3%) | 44.1% (38.6%–49.6%) | **Intervals do not overlap** | 56.5% (54.9%–58.3%) | | New frozen encoders | | | | | | [REVE Base](https://bci.report/methods/reve/) — Input: multi-channel montage | 37.5% (33.3%–41.9%) | 25.5% (22.4%–28.8%) | −12.0 pp **Intervals do not overlap** | 74.6% (72.9%–76.4%) | | [REVE Large](https://bci.report/methods/reve/) — Input: multi-channel montage | 52.7% (47.7%–57.8%) | 46.4% (41.7%–51.2%) | −6.4 pp — Intervals overlap | 76.6% (75.2%–78.0%) | | [LUNA Base](https://bci.report/methods/luna/) — Input: multi-channel montage | 8.6% (7.7%–9.5%) | 9.0% (8.1%–10.1%) | +0.5 pp — Intervals overlap | 67.5% (65.3%–69.5%) | | [LUNA Large](https://bci.report/methods/luna/) — Input: multi-channel montage | 10.0% (8.9%–11.1%) | 11.2% (9.8%–12.6%) | +1.2 pp — Intervals overlap | 72.4% (70.4%–74.0%) | | [BrainOmni Base](https://bci.report/methods/brainomni/) — Input: multi-channel montage | 21.7% (18.6%–25.0%) | 13.0% (10.6%–15.6%) | −8.8 pp **Intervals do not overlap** | 67.7% (65.8%–69.5%) | | [CodeBrain](https://bci.report/methods/codebrain/) — Input: multi-channel montage | 11.9% (10.5%–13.3%) | 11.2% (9.8%–12.7%) | −0.7 pp — Intervals overlap | 70.7% (68.8%–72.7%) | | [EEGMamba](https://bci.report/methods/eegmamba/) — Input: multi-channel montage | 40.2% (35.7%–44.7%) | 32.7% (28.9%–36.7%) | −7.5 pp — Intervals overlap | 64.1% (61.6%–66.4%) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) — Input: multi-channel montage | 17.1% (15.5%–18.8%) in the authors' published pretraining list | 17.5% (15.8%–19.3%) in the authors' published pretraining list | +0.4 pp — Intervals overlap | 77.3% (75.6%–78.8%) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) — Input: multi-channel montage | 22.6% (20.1%–25.3%) in the authors' published pretraining list | 22.4% (19.9%–25.0%) in the authors' published pretraining list | −0.2 pp — Intervals overlap | 79.6% (78.2%–81.0%) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) — Input: multi-channel montage | 54.8% (49.4%–60.0%) | 47.1% (42.1%–52.1%) | −7.7 pp — Intervals overlap | 67.9% (65.4%–70.4%) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) — Input: multi-channel montage | 18.4% (16.3%–20.5%) | 20.9% (18.4%–23.5%) | +2.5 pp — Intervals overlap | 71.0% (69.5%–72.6%) | | [SingLEM](https://bci.report/methods/singlem/) — Input: single-channel | 26.6% (23.7%–29.5%) | 20.5% (17.8%–23.3%) | −6.1 pp **Intervals do not overlap** | 35.3% (33.7%–37.0%) | | [ZUNA 1.1](https://bci.report/methods/zuna/) — Input: channel-wise — Model card: research use only, not for diagnosis or clinical use. | 11.9% (10.8%–13.0%) Exposure unknown | 8.4% (7.5%–9.4%) Exposure unknown | −3.5 pp **Intervals do not overlap** | 60.1% (57.8%–62.3%) Exposure unknown | | EEGPT, frozen · the same runs as the released rows | | | | | | [EEGPT](https://bci.report/methods/eegpt/) | 20.3% (18.2%–22.6%) | 17.3% (15.5%–19.2%) | Intervals overlap | 75.5% (73.9%–77.1%) | **EEGPT on these protocols:** PO5 and PO6 use channel embeddings that EEGPT’s pretraining never trained: they are two of the four names in its 62-name channel dictionary (AF7, AF8, PO5, PO6) that its 58-electrode pretraining montage leaves out. The other six channels here, and all four of the 4-channel subset, use trained embeddings. For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. Eight to four: for a new row, the change and whether the two intervals overlap, as the foundation-model export states them; for a released row and for EEGPT only whether the two intervals overlap, because no file of theirs states that difference. The changes are differences of marginal means, with no paired test. Unless a row is marked, its dataset is not in the model authors’ published pretraining list (checked 2026-10-04; EEGPT’s on 2026-10-12). BETA and sleep · the new frozen encoders ### Small cohorts, descriptive intervals BETA has 70 people and the sleep protocol 20. Each interval is a participant bootstrap of one row on its own: it ignores cross-validation dependence and carries no multiplicity correction, and with 126 frozen cells from the further encoders and EEGPT’s eight more an occasional non-overlap is expected by chance. ### Method comparisons BETA is uniform over its 40 targets and the sleep protocol is a balanced sample of epochs: these rows compare methods; they are not detection, false-alarm or latency estimates for real use. ### New people, same task and setup The folds hold out people. Nothing here is cross-day, cross-device or cross-dataset evidence, and nothing is a clinical claim. ### Short windows Every encoder sees the published 1–2 s windows (sleep as 15 two-second segments), usually shorter than its pretraining context; each model’s input handling is in its row footnote on the protocol pages. ### Software subsets of one recording The eight and four electrodes are channels selected in software from the same 64-channel laboratory recordings. A subset picked this way does not validate a physical low-channel cap, and near-floor scores are not ordered reliably between the two subsets. ### Scalp sleep from EESM19 The six-channel column is the core matrix’s sleep protocol, a balanced, quality-screened scalp sample of EESM19 — not EESM23 above and not ear-EEG. ### One fixed recipe, family order Every new row is the published frozen probe with only the encoder swapped. Rows keep the export’s family order; one row is above another only where their 95% intervals do not overlap. ### Pretraining exposure BETA is in ST-EEGFormer’s published pretraining list, so its BETA cells are marked; ZUNA 1.1 publishes no list, so every ZUNA 1.1 cell is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. The other cells are not in the authors’ published lists (checked 2026-10-04; EEGPT’s on 2026-10-12), which is not proof that the recordings were never seen. ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. The full rows, with footnotes and exposure sources, are on the protocol pages: [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9), [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9), [Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9). The four-minus-eight comparison is in [foundation-models-update.json](https://bci.report/data/foundation-models-update.json). EEGPT’s rows are in their own section of the same protocol pages, from [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json), release [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013). Simulated cap shift · BETA · six decoders ## What if the cap sits one electrode off? A simulation on BETA’s 70 people (40 targets, 2-second windows, chance 2.5%). Six decoders were fitted at the eight posterior electrodes the site uses and then tested with every electrode read one position over: one row forward, or one column to the left or right. The shifted readings are signals recorded at the neighbouring cap sites; nothing was re-recorded. One row forward, training-free CCA changed by −17.6 pp and filter-bank CCA by −14.3 pp; one column to the left, by −5.9 pp and −4.3 pp; to the right, by −1.0 pp and −1.2 pp, both unresolved. eTRCA, calibrated per person at the original positions, fell from 83.5% to 41.2%–55.3%, and EEGNet, trained across people, from 56.0% to 28.6%–40.9%. Recalibrated on the shifted cap, eTRCA reached 77.7%–82.5%. The decoders, shifts, windows and contrasts were prespecified and independently reviewed before any shifted score; an independent audit recomputed every person’s accuracy, the twelve primary intervals and the counts of people from the sealed predictions. The twelve prespecified contrasts: shifted minus original accuracy, paired by person, decoders fitted at the original positions, 2-second windows. Intervals are 99.6% percentile bootstrap intervals over the 70 people (Bonferroni over the twelve), with the fitted models held fixed. | Decoder | Original | Shifted | Change (99.6% interval) | Reading | People losing ≥10 pp, of 70 | | --- | --- | --- | --- | --- | --- | | Forward one row | | | | | | | [CCA](https://bci.report/methods/cca/) | 63.1% | 45.5% | −17.6 pp — (−22.5 to −13.0 pp) | Lower | 49 | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | 79.1% | 64.7% | −14.3 pp — (−20.5 to −8.7 pp) | Lower | 38 | | [eTRCA](https://bci.report/methods/etrca/) | 83.5% | 55.3% | −28.2 pp — (−36.9 to −20.7 pp) | Lower | 54 | | [EEGNet](https://bci.report/methods/eegnet/) | 56.0% | 40.9% | −15.0 pp — (−22.2 to −8.0 pp) | Lower | 36 | | Left one column | | | | | | | [CCA](https://bci.report/methods/cca/) | 63.1% | 57.2% | −5.9 pp — (−8.5 to −3.5 pp) | Lower | 19 | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | 79.1% | 74.7% | −4.3 pp — (−6.7 to −2.2 pp) | Lower | 12 | | [eTRCA](https://bci.report/methods/etrca/) | 83.5% | 46.5% | −36.9 pp — (−45.9 to −28.6 pp) | Lower | 58 | | [EEGNet](https://bci.report/methods/eegnet/) | 56.0% | 30.4% | −25.6 pp — (−33.2 to −18.3 pp) | Lower | 50 | | Right one column | | | | | | | [CCA](https://bci.report/methods/cca/) | 63.1% | 62.1% | −1.0 pp — (−2.9 to +0.7 pp) | Unresolved | 4 | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | 79.1% | 77.9% | −1.2 pp — (−2.7 to +0.3 pp) | Unresolved | 2 | | [eTRCA](https://bci.report/methods/etrca/) | 83.5% | 41.2% | −42.2 pp — (−50.8 to −34.2 pp) | Lower | 64 | | [EEGNet](https://bci.report/methods/eegnet/) | 56.0% | 28.6% | −27.4 pp — (−36.0 to −18.8 pp) | Lower | 49 | People losing 10 pp or more are counted from each person’s trials, as a description; a change of exactly 10 pp (16 of 160 trials; 48 of 480 for EEGNet, which pools three training seeds) counts. Mean accuracy over the 70 people, 2-second windows. | Decoder | Training | Original | Forward one row | Left one column | Right one column | | --- | --- | --- | --- | --- | --- | | Deployed: fitted at the original positions | | | | | | | [CCA](https://bci.report/methods/cca/) | none | 63.1% | 45.5% | 57.2% | 62.1% | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | none | 79.1% | 64.7% | 74.7% | 77.9% | | [eTRCA](https://bci.report/methods/etrca/) | per person, leave one block out (three training trials per target) | 83.5% | 55.3% | 46.5% | 41.2% | | Spectral ridge | across people, seven participant folds | 50.7% | 37.4% | 40.2% | 40.4% | | [EEGNet](https://bci.report/methods/eegnet/) | across people, seven participant folds, from scratch; three training seeds averaged | 56.0% | 40.9% | 30.4% | 28.6% | | [CBraMod](https://bci.report/methods/cbramod/) | frozen encoder, ridge readout across people, seven participant folds | 33.7% | 24.3% | 18.4% | 17.5% | | Recalibrated on the shifted cap | | | | | | | [eTRCA](https://bci.report/methods/etrca/) | per person, leave one block out (three training trials per target) | — | 77.7% | 79.3% | 82.5% | | Spectral ridge | across people, seven participant folds | — | 45.6% | 47.0% | 49.8% | | [EEGNet](https://bci.report/methods/eegnet/) | across people, seven participant folds, from scratch; three training seeds averaged | — | 42.1% | 49.7% | 55.1% | | [CBraMod](https://bci.report/methods/cbramod/) | frozen encoder, ridge readout across people, seven participant folds | — | 28.9% | 29.2% | 35.2% | The neighbour maps: which recorded sites stand in for the eight original ones. | Shift | Sites read, in the order PO5 PO3 POZ PO4 PO6 O1 OZ O2 | Original sites kept, of eight | Mean read-site displacement | | --- | --- | --- | --- | | forward one row | P3 P1 PZ P2 P4 PO3 POZ PO4 | 3 | 3.48 cm | | left one column | P5 PO5 PO3 POZ O2 PO7 O1 OZ | 6 | 2.82 cm | | right one column | O1 POZ PO4 PO6 P6 OZ O2 PO8 | 6 | 2.82 cm | Without the three people flagged by a label-free channel check made before decoding (67 people), every primary mean moves by at most 1.1 pp and keeps its sign. At the original positions CCA (63.1%), spectral ridge (50.7%) and CBraMod (33.7%) reproduce their published two-second, eight-electrode cells exactly, and filter-bank CCA (79.1%) equals its [core-matrix cell](https://bci.report/protocols/beta-8ch/#core-rows); EEGNet’s mean over three training seeds is 56.0% beside the published one-seed cell, 55.8%, which ran on another device. Secondary results, each with a nominal 95.0% interval The secondary results carry no correction across contrasts; their intervals are nominal and support no claim across the family. ### Spectral ridge and CBraMod: shifted minus original | Decoder | Forward one row | Left one column | Right one column | | --- | --- | --- | --- | | Spectral ridge | −13.3 pp — (−16.8 to −10.0 pp) | −10.5 pp — (−12.7 to −8.3 pp) | −10.3 pp — (−13.3 to −7.5 pp) | | [CBraMod](https://bci.report/methods/cbramod/) | −9.3 pp — (−12.7 to −5.9 pp) | −15.2 pp — (−17.8 to −12.8 pp) | −16.2 pp — (−19.6 to −13.0 pp) | ### Mismatch: deployed minus recalibrated | Decoder | Forward one row | Left one column | Right one column | | --- | --- | --- | --- | | [eTRCA](https://bci.report/methods/etrca/) | −22.5 pp — (−28.8 to −16.7 pp) | −32.7 pp — (−38.8 to −27.0 pp) | −41.3 pp — (−47.1 to −35.7 pp) | | Spectral ridge | −8.2 pp — (−10.4 to −6.0 pp) | −6.8 pp — (−8.9 to −4.7 pp) | −9.4 pp — (−12.2 to −6.7 pp) | | [EEGNet](https://bci.report/methods/eegnet/) | −1.2 pp — (−3.6 to +1.4 pp) | −19.3 pp — (−24.0 to −14.8 pp) | −26.4 pp — (−32.0 to −21.0 pp) | | [CBraMod](https://bci.report/methods/cbramod/) | −4.5 pp — (−6.3 to −2.7 pp) | −10.7 pp — (−12.8 to −8.8 pp) | −17.8 pp — (−21.6 to −14.1 pp) | Recalibration uses the person’s own shifted trials (eTRCA) or other people recorded with the same shift (ridge, EEGNet, CBraMod); both assume the shift is known in advance, so they show what was lost, not a practical fix. ### Loss remaining after recalibrating | Decoder | Forward one row | Left one column | Right one column | | --- | --- | --- | --- | | [eTRCA](https://bci.report/methods/etrca/) | −5.8 pp — (−8.3 to −3.2 pp) | −4.2 pp — (−5.5 to −3.0 pp) | −0.9 pp — (−1.8 to −0.1 pp) | | Spectral ridge | −5.1 pp — (−7.5 to −2.8 pp) | −3.7 pp — (−4.9 to −2.5 pp) | −0.9 pp — (−2.0 to +0.3 pp) | | [EEGNet](https://bci.report/methods/eegnet/) | −13.9 pp — (−18.7 to −9.1 pp) | −6.3 pp — (−8.5 to −4.1 pp) | −0.9 pp — (−2.9 to +1.0 pp) | | [CBraMod](https://bci.report/methods/cbramod/) | −4.8 pp — (−7.2 to −2.5 pp) | −4.5 pp — (−6.0 to −3.0 pp) | +1.6 pp — (+0.2 to +3.0 pp) | ### Spline doses: interpolated cap minus original | Decoder | 1 cm | 2 cm | 3 cm | | --- | --- | --- | --- | | Forward one row | | | | | [CCA](https://bci.report/methods/cca/) | −1.9 pp — (−2.8 to −1.0 pp) | −6.7 pp — (−8.3 to −5.1 pp) | −14.9 pp — (−17.7 to −12.1 pp) | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | −0.9 pp — (−1.7 to −0.1 pp) | −4.3 pp — (−6.0 to −2.7 pp) | −11.9 pp — (−15.6 to −8.3 pp) | | [eTRCA](https://bci.report/methods/etrca/) | −4.8 pp — (−7.8 to −2.6 pp) | −12.1 pp — (−16.9 to −8.0 pp) | −25.2 pp — (−31.1 to −19.8 pp) | | Spectral ridge | −1.7 pp — (−3.3 to −0.04 pp) | −4.7 pp — (−7.2 to −2.2 pp) | −8.9 pp — (−12.0 to −5.9 pp) | | [EEGNet](https://bci.report/methods/eegnet/) | −1.1 pp — (−3.1 to +0.7 pp) | −5.7 pp — (−8.9 to −2.5 pp) | −12.2 pp — (−16.7 to −7.8 pp) | | [CBraMod](https://bci.report/methods/cbramod/) | −1.7 pp — (−2.7 to −0.6 pp) | −3.5 pp — (−5.5 to −1.5 pp) | −8.3 pp — (−11.3 to −5.3 pp) | | Left one column | | | | | [CCA](https://bci.report/methods/cca/) | −1.0 pp — (−1.6 to −0.4 pp) | −3.5 pp — (−4.9 to −2.3 pp) | −7.5 pp — (−9.4 to −5.7 pp) | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | −0.8 pp — (−1.3 to −0.3 pp) | −2.7 pp — (−3.7 to −1.8 pp) | −5.6 pp — (−7.2 to −4.0 pp) | | [eTRCA](https://bci.report/methods/etrca/) | −5.8 pp — (−8.2 to −3.8 pp) | −18.5 pp — (−22.7 to −14.5 pp) | −38.7 pp — (−44.6 to −33.0 pp) | | Spectral ridge | −1.6 pp — (−2.6 to −0.5 pp) | −5.1 pp — (−6.8 to −3.3 pp) | −10.2 pp — (−12.3 to −8.2 pp) | | [EEGNet](https://bci.report/methods/eegnet/) | −3.5 pp — (−5.5 to −1.5 pp) | −12.5 pp — (−16.6 to −8.6 pp) | −26.2 pp — (−31.8 to −20.8 pp) | | [CBraMod](https://bci.report/methods/cbramod/) | −2.5 pp — (−3.7 to −1.4 pp) | −8.7 pp — (−10.8 to −6.6 pp) | −15.0 pp — (−17.9 to −12.3 pp) | | Right one column | | | | | [CCA](https://bci.report/methods/cca/) | −0.4 pp — (−1.2 to +0.3 pp) | −0.6 pp — (−1.7 to +0.6 pp) | −1.5 pp — (−2.8 to −0.3 pp) | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | −0.2 pp — (−0.8 to +0.3 pp) | −0.6 pp — (−1.5 to +0.4 pp) | −1.6 pp — (−2.7 to −0.5 pp) | | [eTRCA](https://bci.report/methods/etrca/) | −4.5 pp — (−6.3 to −2.9 pp) | −16.0 pp — (−19.7 to −12.5 pp) | −38.7 pp — (−44.6 to −32.8 pp) | | Spectral ridge | −0.8 pp — (−1.8 to +0.2 pp) | −4.1 pp — (−6.1 to −2.4 pp) | −10.7 pp — (−13.4 to −8.1 pp) | | [EEGNet](https://bci.report/methods/eegnet/) | −2.8 pp — (−4.9 to −0.8 pp) | −14.7 pp — (−19.4 to −10.4 pp) | −31.7 pp — (−37.8 to −25.7 pp) | | [CBraMod](https://bci.report/methods/cbramod/) | −2.3 pp — (−3.4 to −1.2 pp) | −8.7 pp — (−11.1 to −6.4 pp) | −16.6 pp — (−20.0 to −13.3 pp) | An interpolated site is a weighted mix of recorded sites, so for decoders fitted to positions the spline doses mix displacement with interpolation and are indicative only. Rebuilding each original site from the other sites with no shift changes CCA by −1.7 pp, Filter-bank CCA by −1.2 pp, eTRCA by −22.6 pp, Spectral ridge by −2.6 pp, EEGNet by −14.5 pp and CBraMod by −11.7 pp. At the neighbour maps’ angles the spline gave eTRCA +10.2 pp more than the right-hand map and +3.8 pp more on the left. ### 1-second windows: shifted minus original | Decoder | Forward one row | Left one column | Right one column | | --- | --- | --- | --- | | [CCA](https://bci.report/methods/cca/) | −8.9 pp — (−11.1 to −6.8 pp) | −3.0 pp — (−4.4 to −1.7 pp) | −0.3 pp — (−1.4 to +0.9 pp) | | [Filter-bank CCA](https://bci.report/methods/fbcca/) | −15.2 pp — (−18.1 to −12.4 pp) | −7.1 pp — (−8.5 to −5.6 pp) | −1.2 pp — (−2.7 to +0.2 pp) | | [eTRCA](https://bci.report/methods/etrca/) | −31.9 pp — (−37.1 to −27.0 pp) | −38.7 pp — (−43.9 to −33.6 pp) | −44.5 pp — (−49.8 to −39.3 pp) | ### Differences in drop: the first decoder’s change minus the second’s, same shift | Decoder | Forward one row | Left one column | Right one column | | --- | --- | --- | --- | | eTRCA − Filter-bank CCA | −13.9 pp — (−21.0 to −7.1 pp) | −32.6 pp — (−38.4 to −27.0 pp) | −41.1 pp — (−46.8 to −35.5 pp) | | EEGNet − CCA | +2.6 pp — (−1.0 to +6.2 pp) | −19.7 pp — (−24.6 to −15.0 pp) | −26.3 pp — (−32.1 to −20.6 pp) | Limits of the simulation The simulation moves the whole cap as one piece with the reference fixed; a real slip can be partial, change contact and impedance, and move the reference too. Backward slips and slides toward one ear are not simulated. BETA gives template positions, so each person’s true displacement is unknown. The neighbour maps move the read sites a mean of 3.48 cm forward and 2.82 cm sideways and keep 3 and 6 of the eight original sites, so the directions are not compared with each other. Intervals resample people with the fitted models held fixed; the variance of training is not included. One dataset and offline 40-target classification, with no latency or online ITR measured; nothing follows about any headset. The decoders start from very different accuracies, so a change in points is read beside its baseline. ### Data and method credits BETA · Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · participant consent described non-commercial scientific research · [Bingchuan/BETA ↗](https://huggingface.co/datasets/Bingchuan/BETA) CBraMod, frozen: code MIT, weights Apache-2.0 as its official model card states; not distributed. [Method page](https://bci.report/methods/cbramod/) Nakanishi et al. 2018 (ensemble TRCA) · Chen et al. 2015, doi:10.1088/1741-2560/12/4/046008 (filter-bank CCA) Data source: [reviewed aggregate JSON](https://bci.report/data/electrode-shift-update.json) · schema bci-report-electrode-shift-update-v1. Methods & limits ## Limits of the two paired comparisons Sleep · in-ear and scalp ### Six scalp electrodes, not eight The scalp files also carry two mastoid slots. They were used only in the completeness screen and never entered the model. ### Position, count and hardware move together The in-ear and scalp configurations differ in placement, geometry and number of electrodes. The difference is descriptive, not an estimate of what the ear hardware alone costs. ### One simple, untuned model Five log-bandpower features per channel and an L2 logistic regression with no hyperparameter search; a stronger model could score higher on either configuration. Eyes open and closed · four and sixteen ### A pipeline check Closing the eyes produces a strong alpha rhythm over the back of the head, which is why four posterior electrodes can carry this task. Concentration, workload and intent detection each need their own test. ### Chosen in software Four channels selected in software from the 16-electrode wet-cap recordings do not validate a physical four-channel device. ### 19 of 20 recordings The source has 20 recordings. The authors’ own data loader leaves out recording 07, and this comparison keeps that fixed cohort. Our event audit also flagged overlapping events in that recording, but the authors’ reason for leaving it out is not established. Filter-bank CCA, CCA and learned heads on CCA’s statistics are also compared with fewer of BETA’s posterior electrodes and with shorter windows, on SSVEP frequencies held out of training: [What lets a decoder answer an SSVEP frequency it never trained on?](https://bci.report/topics/new-ssvep-frequencies/#practical-settings) Two frozen encoders and a log band-power ridge are also read with some channels zeroed (channels zeroed after preprocessing, not a physical montage), with the head kept or refitted: [Channels zeroed on two published encoders, head kept or refitted →](https://bci.report/topics/pilot-studies/#nsp-components-a) ### EESM23 · Ear-EEG Sleep Monitoring 2023 Dataset: Yousef Rezaei Tabar, Kaare Mikkelsen, Laura Birch, Nelly Shenton, Simon L. Kappel, Astrid R. Bertelsen, Reza Nikbakht, Hans O. Toft, Chris H. Henriksen, Martin C. Hemmsen, Mike L. Rank, Marit Otto and Preben Kidmose · Ear-EEG Sleep Monitoring 2023 (EESM23), OpenNeuro ds005178 v1.0.0, doi:10.18112/openneuro.ds005178.v1.0.0. Study: Kaare Bjarke Mikkelsen, Yousef Rezai Tabar, Laura Rævsbæk Birch, Simon Lind Kappel, Christian Bech Christensen, Lars Dalskov Mosgaard, Marit Otto, Martin Christian Hemmsen, Mike Lind Rank and Preben Kidmose · Ear-EEG sleep monitoring data sets, Scientific Data 12, 301 (2025), doi:10.1038/s41597-025-04579-8. Processed export: Zachary1150/EESM23-Processed on Hugging Face, which did not create the original cohort. Consent covered the study; the data were released after the GDPR office of Region Midt judged them fully anonymized. Only cohort aggregates appear here. [OpenNeuro dataset ↗](https://doi.org/10.18112/openneuro.ds005178.v1.0.0) · [Scientific Data paper ↗](https://doi.org/10.1038/s41597-025-04579-8) · [Processed export ↗](https://huggingface.co/datasets/Zachary1150/EESM23-Processed/tree/aed9526a1a56c6644b13a8ef95e03e6a6729a87f) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) ### EEG Alpha Waves dataset Grégoire Cattan, Pedro L. C. Rodrigues and Marco Congedo · EEG Alpha Waves dataset, Zenodo, doi:10.5281/zenodo.2605110. Report: Grégoire Cattan, Pedro Luiz Coelho Rodrigues and Marco Congedo · EEG Alpha Waves Dataset, research report, GIPSA-lab, Grenoble, 2018, hal-02086581. According to the primary report, every participant gave written informed consent covering the experimental process, the data management procedures and the right to withdraw at any moment. The report names no ethics committee. Only cohort aggregates appear here. [Zenodo record ↗](https://doi.org/10.5281/zenodo.2605110) · [HAL report ↗](https://hal.science/hal-02086581/document) · [Authors’ code ↗](https://github.com/plcrodrigues/py.ALPHA.EEG.2017-GIPSA) · [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/evidence-update.json) · schema bci-report-evidence-update-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Can fewer electrodes, or electrodes in the ear, match a full scalp montage?* https://bci.report/topics/fewer-electrodes/ Figures from releases [`electrode-shift-update-20261106`](https://bci.report/releases/#electrode-shift-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`evidence-update-20260922`](https://bci.report/releases/#evidence-update-20260922) (2026-09-22), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/fewer-electrodes/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Does EEG decoding still work while walking or running? Motion robustness · scalp and ear EEG Performance declines as participants move. - [Reviewed aggregate JSON ↓](https://bci.report/data/deployment-topics.json) - Measured on: [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/); [Phantom EEG with motion, muscle and eye artifacts](https://bci.report/datasets/phantom/); [Multimodal gait dataset · treadmill walking](https://bci.report/datasets/gait-eeg/) · Methods: [LaBraM](https://bci.report/methods/labram/); [CBraMod](https://bci.report/methods/cbramod/); [CCA](https://bci.report/methods/cca/) ## Short answer Less well. With the same **23** people on a three-target, two-second SSVEP task (eight selected scalp channels; chance **33.3%**), the training-free author-style CCA reference scored **82.3%** balanced accuracy standing, **70.7%** at a slow walk and **68.6%** at a fast walk. On an ERP task, with the same **17** people in both conditions, a decoder fit on each person's first standing session and then used unchanged saw its scalp ROC AUC change by **−0.276 AUC** from standing to running (paired 95% interval **−0.322** to **−0.227 AUC**). Both tasks are derivatives of one original mobile study. **Author-style CCA, the same 23 people standing and walking.** Participant-mean balanced accuracy of the training-free reference on three-target, two-second SSVEP with eight selected scalp channels. Dot: the estimate; line: participant bootstrap 95% interval; dashed line: chance, 33.3%. - Standing: 82.3% (74.6%–89.4%) - Slow walk · 0.8 m/s: 70.7% (62.3%–78.6%) - Fast walk · 1.6 m/s: 68.6% (60.1%–76.7%) Matched operating point ## Two seconds, eight selected scalp channels All four methods use the window [0, 2 s) and the same eight posterior scalp channels, and are evaluated on 23 people with three SSVEP targets. Participant-mean balanced accuracy; participant bootstrap 95% intervals are shown below each score. | Method | Standing | Slow walk · 0.8 m/s | Fast walk · 1.6 m/s | | --- | --- | --- | --- | | Author-style CCA — analytic reference | 82.3% (74.6%–89.4%) | 70.7% (62.3%–78.6%) | 68.6% (60.1%–76.7%) | | Spectral ridge — trained on the other 22 standing participants | 78.7% (70.7%–86.1%) | 69.9% (61.2%–78.1%) | 68.6% (60.4%–76.3%) | | LaBraM — frozen encoder + ridge, trained on other participants | 59.5% (53.6%–65.5%) | 54.7% (49.5%–60.1%) | 47.8% (43.3%–52.3%) | | CBraMod — frozen encoder + ridge, trained on other participants | 59.9% (54.3%–65.4%) | 46.2% (41.1%–51.4%) | 46.0% (42.2%–49.6%) | Bars span 0–100%; chance is 33.3%. LaBraM uses verified standard channel IDs; CBraMod uses an exploratory positional channel order. Record-level pretraining exposure is not certified. A separate operating point ## Five-second scalp and ear comparisons Here the two classical methods use a five-second window instead of two. Participant-mean balanced accuracy · 23 people · 3 targets. Scalp uses 8 selected channels from 32 recorded; ear uses 14 contacts. | Method / modality | Standing | Slow walk · 0.8 m/s | Fast walk · 1.6 m/s | | --- | --- | --- | --- | | Author-style CCA — 8 selected scalp channels · 5 s | 88.7% (80.0%–95.9%) | 82.9% (74.3%–90.4%) | 80.9% (72.0%–88.8%) | | Spectral ridge — 8 selected scalp channels · 5 s | 85.0% (77.0%–92.0%) | 79.2% (70.1%–87.5%) | 77.2% (67.7%–85.7%) | | Author-style CCA — 14 ear contacts · 5 s | 53.1% (47.4%–59.6%) | 43.4% (39.6%–47.8%) | 39.1% (36.7%–41.8%) | | Spectral ridge — 14 ear contacts · 5 s | 52.0% (46.9%–57.8%) | 45.4% (40.7%–51.2%) | 40.0% (36.4%–43.7%) | ERP session transfer ## AUC falls toward chance while moving For each participant and modality, a temporal-feature L2 logistic regression is fit on the first standing session and then used unchanged. The endpoint is ROC AUC. **−0.276 AUC** Scalp ROC AUC change: standing to running, same 17 people ### The running contrast is paired on the same people Running has 17 participants, standing and walking 24. Among the same 17 people, the standing-to-running change is −0.276 AUC for scalp (paired 95% interval −0.322 to −0.227 AUC) and −0.225 AUC for ear (−0.277 to −0.175 AUC). Participant-mean ROC AUC. Scalp uses 32 channels; ear uses 14 contacts. Values in parentheses are participant bootstrap 95% intervals. | Modality | Standing | Slow walk · 0.8 m/s | Fast walk · 1.6 m/s | Running · 2.0 m/s | | --- | --- | --- | --- | --- | | Scalp · 32 channels — standing-session fit · unchanged weights | 0.859 (0.820–0.897) — n=24 | 0.740 (0.703–0.777) — n=24 | 0.635 (0.602–0.673) — n=24 | 0.586 (0.561–0.613) — n=17 | | Ear · 14 contacts — standing-session fit · unchanged weights | 0.734 (0.694–0.774) — n=24 | 0.628 (0.595–0.661) — n=24 | 0.563 (0.540–0.587) — n=24 | 0.517 (0.491–0.541) — n=17 | Engineering control · physical phantom ## Artifacts, measured without people A conductive head phantom broadcast ten known brain sources while artifacts were added one kind at a time. A linear decoder trained on the clean recording was tested on held-out time blocks of every condition. One phantom is one experimental unit. **Neither column is accuracy.** Correlation asks whether the recovered source moves with the true one. Predictive R² also scores amplitude and has no floor: below zero, the prediction is further from the true source than a flat line would be. Median across five folds of the within-fold median across ten brain sources. The same 60-second blocks are held out in every condition, with a two-second purge on each side of a training block. Trained on the brain-only recording. | Condition | Signed correlation r | Predictive R² | | --- | --- | --- | | Brain sources only — training condition, held-out time | 0.761 (fold range 0.755 to 0.772) | 0.567 (fold range 0.549 to 0.594) | | Eye artifacts | 0.492 (fold range 0.432 to 0.569) | −33.612 (fold range −75.205 to −23.971) | | Walking | 0.549 (fold range 0.542 to 0.577) | −795.772 (fold range −935.028 to −425.205) | | Neck muscle | 0.310 (fold range 0.249 to 0.317) | −596.820 (fold range −1106.382 to −281.754) | | Facial muscle | 0.143 (fold range 0.127 to 0.155) | −164.313 (fold range −270.004 to −110.197) | | All artifacts together | 0.115 (fold range 0.104 to 0.131) | −142.832 (fold range −172.296 to −100.142) | ### Phantom EEG with motion, muscle and eye artifacts Ryan J. Downey and Daniel P. Ferris · Phantom EEG Dataset with Motion, Muscle, and Eye Artifacts and Example Scripts, OpenNeuro ds004784 v1.0.4, doi:10.18112/openneuro.ds004784.v1.0.4. Method: doi:10.3390/s23198214. One electrically conductive head phantom; no human participants. [OpenNeuro dataset ↗](https://doi.org/10.18112/openneuro.ds004784.v1.0.4) · [Method paper ↗](https://doi.org/10.3390/s23198214) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/evidence-update.json) · schema bci-report-evidence-update-v1. Treadmill speed · dry electrodes ## What a walking-speed score may be measuring 58 people walked on a treadmill at three speeds while a dry-electrode headset recorded 19 scalp channels. A fixed classifier guessed the speed from each whole 58-second recording. A second classifier saw only three movement-nuisance features: amplitude, low-frequency and line-frequency power. Chance is 33.3%. **+5.2 pp** Nuisance features minus spectral bands, same 58 people ### Nuisance features alone score in the same range as the spectral bands 50.6% against 45.4%, with overlapping intervals (44.3%–56.9% and 39.1%–51.7%). A walking-speed score can be reached without any spectral pattern, so a score like this alone is not evidence of decoding the brain. Participant-held-out balanced accuracy over five folds, with its participant-bootstrap 95% interval, and macro F1. Every person contributes all three speeds. | Input | Balanced accuracy | Macro F1 | | --- | --- | --- | | Relative spectral bands — five relative bands on each of 19 scalp channels | 45.4% (39.1%–51.7%) | 45.5% | | Movement-nuisance features — comparator · three features, no spectral pattern | 50.6% (44.3%–56.9%) | 48.6% | ### No motion sensors in the model The source also recorded inertial and muscle signals; they were not used as controls. Speed order was not verified to be counterbalanced, so fatigue, drift and electrode contact may change with it. ### 1 person excluded before scoring One source file has malformed channel labels, so that person's 3 recordings were excluded; no channel was relabelled. ### Multimodal gait · EEG during treadmill walking at three speeds R. Katmah, A. AlShehhi, D. Kosaji, N. Al-Rahmani, M. Abdullah, A. A. V. Hulleck and K. Khalaf · A multimodal gait dataset of brain activity, muscle activity, kinematics and ground forces in young adults, PhysioNet v1.0.0 (2026), doi:10.13026/r0ea-7161. The study followed the Declaration of Helsinki and was approved by the Institutional Review Board of Khalifa University (H19-038); all participants gave written informed consent. Only cohort aggregates appear here. [PhysioNet dataset ↗](https://doi.org/10.13026/r0ea-7161) · [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/context-update.json) · schema bci-report-context-update-v1. Methods & limits ## Limits ### Scalp 8 is a selection The mobile SSVEP scalp view selects eight posterior channels from 32 recorded; the dry-versus-wet page uses a native eight-channel headband. ### Balanced accuracy differs from AUC The SSVEP endpoint is participant-mean balanced accuracy for a three-class task. The ERP endpoint is participant-mean ROC AUC. The two values are on different scales. ### The derivatives share a study NEMAR nm000125 and nm000201 are SSVEP and ERP derivatives of the same original mobile BCI study, with different assessed sample counts. ### Lee2021 – SSVEP paradigm of the Mobile BCI dataset Lee, Y., Shin, G., Lee, M., & Lee, S. (2026). Lee2021 – SSVEP paradigm of the Mobile BCI dataset (Version v1.0.2) [Data set]. NEMAR. https://doi.org/10.82901/nemar.nm000125 Lee, Y.-E., Shin, G.-H., Lee, M., & Lee, S.-W. (2021). Mobile BCI dataset of scalp- and ear-EEGs with ERP and SSVEP paradigms while standing, walking, and running. Scientific Data, 8, 315. https://doi.org/10.1038/s41597-021-01094-4 NEMAR/MOABB-generated BIDS derivative · CC BY 4.0. [SSVEP dataset ↗](https://nemar.org/dataset/nm000125) · [Source study ↗](https://doi.org/10.1038/s41597-021-01094-4) ### ERP paradigm of the Mobile BCI dataset Lee, Y., Shin, G., Lee, M., & Lee, S. (2026). ERP paradigm of the Mobile BCI dataset (Version v1.0.2) [Data set]. NEMAR. https://doi.org/10.82901/nemar.nm000201 Lee, Y.-E., Shin, G.-H., Lee, M., & Lee, S.-W. (2021). Mobile BCI dataset of scalp- and ear-EEGs with ERP and SSVEP paradigms while standing, walking, and running. Scientific Data, 8, 315. https://doi.org/10.1038/s41597-021-01094-4 NEMAR/MOABB-generated BIDS derivative · CC BY 4.0. [ERP dataset ↗](https://nemar.org/dataset/nm000201) · [Source study ↗](https://doi.org/10.1038/s41597-021-01094-4) Data source: [reviewed aggregate JSON](https://bci.report/data/deployment-topics.json) · schema bci-report-public-deployment-topics-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Does EEG decoding still work while walking or running?* https://bci.report/topics/on-the-move/ Figures from releases [`context-update-20260927`](https://bci.report/releases/#context-update-20260927) (2026-09-27), [`evidence-update-20260922`](https://bci.report/releases/#evidence-update-20260922) (2026-09-22), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/on-the-move/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Does a decoder trained on an earlier session still work later? Session transfer · the same person, no labels from the later session Motor imagery two sessions later: frozen CBraMod with a readout fitted on the earlier session scores above a spectral baseline in both cohorts. Target detection at later visits: a first-visit decoder ranks targets less well at the last nominal visit, and its average precision is lower. Continuous cursor tracking: a fixed decoder does worse than a constant baseline for every admitted record. - [Reviewed aggregate JSON ↓](https://bci.report/data/later-sessions-update.json) - Measured on: [WBCIC-SHU motor imagery dataset](https://bci.report/datasets/wbcic-shu/); [Longitudinal ERP dataset (RSVP)](https://bci.report/datasets/longitudinal-rsvp/); [Forenzo & He continuous-tracking EEG-BCI dataset](https://bci.report/datasets/forenzo-continuous-tracking/) · Methods: [CBraMod](https://bci.report/methods/cbramod/) ## Short answer Partly. On WBCIC-SHU motor imagery — trained on each person’s recording session 1 and tested on session 3 with no labels from it, **51** people in a two-class and **11** in a three-class cohort, kept apart — frozen CBraMod with a ridge readout fitted on session 1 reached **67.7%** and **51.6%** balanced accuracy, **+13.8 pp** and **+14.0 pp** above a relative spectral ridge, each paired interval above zero. This is same-person transfer, not decoding for someone new, and it does not show that pretraining caused the gain. On a longitudinal RSVP task, a decoder trained at each of **15** people’s first visit ranked targets with an AUROC of **0.888** at the publisher’s nominal Day 7 visit and **0.853** at nominal Day 200, a paired change of **−0.035** (**−0.065** to **−0.006**). With targets only about **2.5%** of events, average precision fell from **0.359** to **0.240**. In continuous cursor tracking, a fixed spectral ridge trained on each record’s earliest session had a higher overall error in its latest session than a constant source-mean comparator, for every admitted record in both cohorts and both response arms: a negative result, with a severe upper tail. WBCIC-SHU · motor imagery · session 1 to session 3 ## Motor imagery two sessions later: two baselines and a frozen encoder People imagined moving the left or the right hand (a two-class cohort, 51 people) or, in a separate three-class cohort (11 people), a foot movement as well, in three recording sessions. Each person’s complete session 1 trains the decoder and session 3 is the test; session 2 is unused, and no label from session 3 is used. Session numbers are recording-session ordinals, not a guaranteed time gap or distinct calendar days. The cohorts are kept apart and never averaged. Participant-equal mean balanced accuracy on session 3, with a paired whole-participant bootstrap 95% interval. Chance is 50.0% for two classes and 33.3% for three, where the source-majority prior sits by construction. Accuracy and macro F1 are the secondary means. | Decoder | Two-class motor imagery · 51 people — left or right hand | Three-class motor imagery · 11 people — left hand, right hand or foot | | --- | --- | --- | | Source-majority prior — the floor: predicts session 1’s most frequent class | 50.0% (Accuracy 50.0% · macro F1 33.3%) | 33.3% (Accuracy 33.3% · macro F1 16.7%) | | Relative spectral power + ridge — a fixed CPU baseline | 53.9% (95% interval 52.7%–55.1%) (Accuracy 53.9% · macro F1 49.7%) | 37.6% (95% interval 35.4%–40.0%) (Accuracy 37.6% · macro F1 34.1%) | | CBraMod · frozen encoder, session-1 ridge readout — a pretrained EEG foundation model | 67.7% (95% interval 64.8%–70.5%) (Accuracy 67.7% · macro F1 66.9%) | 51.6% (95% interval 45.2%–58.3%) (Accuracy 51.6% · macro F1 50.5%) | Paired differences in balanced accuracy, in percentage points (pp), the same people and test trials. Each interval resamples matched per-person differences within its cohort, conditional on the fixed trained models and predictions. | Decoder | Two-class motor imagery · 51 people | Three-class motor imagery · 11 people | | --- | --- | --- | | Spectral ridge minus the source prior | +3.9 pp (95% interval +2.7 to +5.1 pp) (Interval above zero.) | +4.3 pp (95% interval +2.1 to +6.6 pp) (Interval above zero.) | | Frozen CBraMod minus the spectral ridge | +13.8 pp (95% interval +11.3 to +16.4 pp) (Interval above zero.) | +14.0 pp (95% interval +7.5 to +20.8 pp) (Interval above zero.) | **+13.8 pp** Frozen CBraMod minus the spectral ridge, two-class cohort, 51 people ### Above the spectral baseline in both cohorts Paired interval +11.3 to +16.4 pp. In the small three-class cohort the gap was +14.0 pp (+7.5 to +20.8 pp). The CBraMod arm was frozen after the spectral results existed and reuses their test trials: a comparative follow-up, not a fresh test set. There is no matched random-weight control, the two feature spaces differ in size and preprocessing (the same ridge alpha does not equalize regularization across them), and whether this checkpoint saw WBCIC-SHU in pretraining is not established, so the gap compares the two pipelines; it does not show that pretraining caused it. Trials: 10,199 training and 10,195 test trials in the two-class cohort, 3,299 and 3,300 in the three-class cohort; no person or trial was excluded. 6 of the 186 session records hold fewer trials than the publisher’s nominal count (40,490 delivered against 40,500); 3 of them are in the sessions used and are admitted under a reviewed variable-N rule, with no filling, truncation or performance-based exclusion; the original strict-count holds stay recorded. Why the trials are missing is unknown. Longitudinal RSVP · target detection · a first-visit decoder ## Does your decoder still work at later visits? Same-person, offline RSVP classification; publisher nominal visit labels; no target-session adaptation. In rapid serial visual presentation (RSVP), a stream of face images is shown and the decoder has to pick out the rare target. 15 people, the source’s Group A, were recorded at visits the publisher labels Day 1, 7, 80 and 200. Each person’s decoder is trained and calibrated on their first visit only, then scored unchanged at each later visit: 72,000 events per visit across the 15 people. Decoder: Normalized ERP features + logistic/Platt (CPU baseline). **AUROC at each nominal later visit: the same people, one first-visit decoder each.** Participant-equal mean AUROC. Dot: the mean; line: its participant-bootstrap 95% interval, printed under each label; dashed line: chance, a decoder that orders events at random. The visits are labels, evenly spaced, not a time axis, and no curve is fitted between them. AUROC ranks targets above non-targets; it is not accuracy. - Nominal Day 7 — 95% interval 0.864–0.911: 0.888 - Nominal Day 80 — 95% interval 0.854–0.903: 0.879 - Nominal Day 200 — 95% interval 0.827–0.879: 0.853 ### Limits of the RSVP result - Nominal visits: Day 7, Day 80 and Day 200 are the publisher’s labels; each person’s actual calendar interval is unverified. - Rare targets: about 2.5% of events at every visit. Average precision is not precision at a chosen threshold, and neither it nor AUROC is an online selection success rate. - Small study: 15 people and one fixed offline baseline. Participant-equal means at each nominal visit, 15 people; counts are event totals. | Measure | Nominal Day 7 | Nominal Day 80 | Nominal Day 200 | | --- | --- | --- | --- | | AUROC — ranks targets, 0 to 1 | 0.888 (95% interval 0.864–0.911) | 0.879 (95% interval 0.854–0.903) | 0.853 (95% interval 0.827–0.879) | | Average precision — 0 to 1 | 0.359 | 0.310 | 0.240 | | Log loss — per event, natural log; lower is better | 0.154 | 0.114 | 0.164 | | Brier score — lower is better | 0.029 | 0.027 | 0.038 | | ECE, 10 bins — expected calibration error; lower is better | 0.023 | 0.020 | 0.039 | | Target events — of all events at the visit | 1,794 of 72,000 (2.5%) | 1,786 of 72,000 (2.5%) | 1,793 of 72,000 (2.5%) | | Delivered events — not available from the protected scoring interface | not available (null, not zero) | not available (null, not zero) | not available (null, not zero) | **−0.035** AUROC, nominal Day 200 minus Day 7, the same 15 people ### AUROC and average precision lower at nominal Day 200 Paired interval −0.065 to −0.006, below zero. 6 of 15 people lost 0.05 AUROC or more. Average precision is lower too: 0.359 at nominal Day 7, 0.240 at Day 200. This describes this cohort and this decoder; it does not show that elapsed time caused the change. Label coding: the paper and the publisher’s sample script code 1 as non-target and 2 as target, and the release’s Trigger.txt reverses them. The paper and script were followed, a precedence decided before modelling. Continuous cursor tracking · a negative result ## A simple baseline exposes cross-session decoder failures Offline, same-person, no target-session adaptation. People used motor imagery to steer a cursor after a target moving continuously across a screen, in two cohorts the publisher calls Main and Transfer Learning. For each admitted record, a fixed spectral ridge is trained on every eligible non-chance run of the earliest complete recorded session and tested on the latest, with no label from it, against a source-mean comparator: the earliest session’s average response, repeated for every row. Cohort records admitted 23 Main 9 of 14, Transfer Learning 14 of 14. Records, not proven unique people. Target rows per response arm 1,999,926 The same rows for both arms: two outcomes do not double the independent sample. No response row was removed. Response fits 46 One per admitted record and response arm; every one succeeded. The error is the joint source-SD-normalized RMSE: errors divided by the earliest session’s response standard deviation, averaged over both axes within each trial before the square root. Lower is better. Each record averages its trials equally, and each cohort mean weights records equally; intervals resample records within a cohort. Two response arms, each with its own table. They are different outcomes on the same rows: never combined, and their errors are not compared with each other. ### Historical decoder velocity (historical decoder imitation) The output of the decoder in use at recording time, as the publisher stored it, imitated: not intended hand motion or intended control. Historical decoder velocity (historical decoder imitation) · error in the latest session, lower is better: the ridge’s mean with its interval and its median, the comparator’s mean, and the paired difference (positive: the ridge’s error is larger). | Cohort | Spectral ridge · mean | Spectral ridge · median | Source-mean comparator · mean | Ridge minus comparator · same records | | --- | --- | --- | --- | --- | | Main cohort — 9 of 14 records admitted | 192.290 (95% interval 1.292 to 574.235) | 1.317 (the mean is 146 times the median) | 1.266 (95% interval 1.214 to 1.316) | 191.024 (95% interval 0.026 to 572.995) (ridge error higher on 9 of 9 records) | | Transfer Learning cohort — 14 of 14 records admitted | 44.172 (95% interval 0.955 to 130.476) | 0.990 (the mean is 45 times the median) | 0.864 (95% interval 0.801 to 0.929) | 43.308 (95% interval 0.105 to 129.601) (ridge error higher on 14 of 14 records) | **Historical decoder velocity: the ridge’s mean and median beside the comparator, on a log scale.** Log scale, powers of ten from 10⁻² to 10⁶: each gridline is a hundredfold step, marking 1, 100 and 10,000. Dot: the value; line: its 95% interval; the ridge’s median has no interval here. - Main · spectral ridge · mean — 95% interval 1.292–574.235: 192.290 - Main · spectral ridge · median: 1.317 - Main · source-mean comparator · mean — 95% interval 1.214–1.316: 1.266 - Transfer Learning · spectral ridge · mean — 95% interval 0.955–130.476: 44.172 - Transfer Learning · spectral ridge · median: 0.990 - Transfer Learning · source-mean comparator · mean — 95% interval 0.801–0.929: 0.864 ### Constructed displacement proxy (constructed proxy) Target position minus cursor position in the publisher’s screen coordinates, a separately fitted proxy. Constructed displacement proxy (constructed proxy) · error in the latest session, lower is better: the ridge’s mean with its interval and its median, the comparator’s mean, and the paired difference (positive: the ridge’s error is larger). | Cohort | Spectral ridge · mean | Spectral ridge · median | Source-mean comparator · mean | Ridge minus comparator · same records | | --- | --- | --- | --- | --- | | Main cohort — 9 of 14 records admitted | 698.499 (95% interval 1.166 to 2,093) | 1.225 (the mean is 570 times the median) | 1.056 (95% interval 0.977 to 1.133) | 697.443 (95% interval 0.121 to 2,092) (ridge error higher on 9 of 9 records) | | Transfer Learning cohort — 14 of 14 records admitted | 69.844 (95% interval 1.071 to 207.287) | 1.182 (the mean is 59 times the median) | 0.916 (95% interval 0.841 to 0.987) | 68.928 (95% interval 0.143 to 206.410) (ridge error higher on 14 of 14 records) | **Constructed displacement proxy: the ridge’s mean and median beside the comparator, on a log scale.** Log scale, powers of ten from 10⁻² to 10⁶: each gridline is a hundredfold step, marking 1, 100 and 10,000. Dot: the value; line: its 95% interval; the ridge’s median has no interval here. - Main · spectral ridge · mean — 95% interval 1.166–2,093: 698.499 - Main · spectral ridge · median: 1.225 - Main · source-mean comparator · mean — 95% interval 0.977–1.133: 1.056 - Transfer Learning · spectral ridge · mean — 95% interval 1.071–207.287: 69.844 - Transfer Learning · spectral ridge · median: 1.182 - Transfer Learning · source-mean comparator · mean — 95% interval 0.841–0.987: 0.916 Across the four cohort-and-arm cells, the ridge’s mean error is 45 to 570 times its median: a few records carry very large errors. No scored record was removed, clipped or winsorized, and what caused the extreme errors — feature distribution shift, scaling sensitivity or something else — is not established. The every-record statement covers the overall primary error only, not the secondary metrics or each recorded-decoder context. Coverage. Main admits 9 of its 14 candidate records and is read conditional on that subset; 5 were held on metadata before any scoring and stay recorded: 2 because a session lacks its required Chance R01 record; 2 because the decoder-local run allocation differs from the documented contract; 1 because the sample-boundary geometry differs; its exact numeric cause was not independently reconstructed. Transfer Learning admits 14 of 14. That is 23 admitted cohort records; the cohorts may share people and are never pooled. Transfer Learning is the publisher’s name for how that cohort’s data were collected; no transfer-learning model was trained here. Secondary metrics per axis — raw RMSE, R² and Pearson correlation Record-equal means with their 95% intervals. Raw RMSE is in each response’s stored units; R² and Pearson are unitless. A constant prediction has no correlation, so the comparator’s Pearson values are undefined: null, not zero. Paired differences for these metrics are not published. Historical decoder velocity · secondary metrics in the latest session. | Cohort and model | Raw RMSE, x | Raw RMSE, y | R², x | R², y | Pearson, x | Pearson, y | | --- | --- | --- | --- | --- | --- | --- | | Main cohort — Spectral ridge | 38.889 (95% interval 0.309 to 116.038) | 50.599 (95% interval 0.312 to 151.159) | −172,983 (95% interval −518,950 to −0.018) | −296,663 (95% interval −889,988 to −0.046) | 0.029 (95% interval 0.00233 to 0.061) | 0.034 (95% interval 0.00105 to 0.073) | | Main cohort — Source-mean comparator | 0.306 (95% interval 0.290 to 0.321) | 0.302 (95% interval 0.289 to 0.313) | −0.000498 (95% interval −0.000786 to −0.000255) | −0.000377 (95% interval −0.000659 to −0.000127) | undefined (null) | undefined (null) | | Transfer Learning cohort — Spectral ridge | 4.443 (95% interval 0.212 to 12.877) | 13.815 (95% interval 0.229 to 40.954) | −5,440 (95% interval −16,319 to −0.188) | −58,117 (95% interval −174,350 to −0.339) | 0.045 (95% interval 0.00527 to 0.096) | 0.019 (95% interval −0.022 to 0.075) | | Transfer Learning cohort — Source-mean comparator | 0.198 (95% interval 0.184 to 0.212) | 0.201 (95% interval 0.187 to 0.217) | −0.000187 (95% interval −0.000314 to −0.0000754) | −0.000937 (95% interval −0.00140 to −0.000531) | undefined (null) | undefined (null) | Constructed displacement proxy · secondary metrics in the latest session. | Cohort and model | Raw RMSE, x | Raw RMSE, y | R², x | R², y | Pearson, x | Pearson, y | | --- | --- | --- | --- | --- | --- | --- | | Main cohort — Spectral ridge | 124.828 (95% interval 0.412 to 373.616) | 332.152 (95% interval 0.431 to 995.534) | −928,350 (95% interval −2,785,050 to −0.183) | −6,442,341 (95% interval −19,327,021 to −0.388) | 0.050 (95% interval 0.031 to 0.074) | 0.072 (95% interval 0.018 to 0.137) | | Main cohort — Source-mean comparator | 0.388 (95% interval 0.366 to 0.413) | 0.376 (95% interval 0.349 to 0.404) | −0.026 (95% interval −0.045 to −0.00771) | −0.021 (95% interval −0.039 to −0.00612) | undefined (null) | undefined (null) | | Transfer Learning cohort — Spectral ridge | 9.772 (95% interval 0.415 to 28.423) | 38.151 (95% interval 0.445 to 113.515) | −21,504 (95% interval −64,511 to −0.318) | −337,640 (95% interval −1,012,920 to −0.391) | 0.035 (95% interval 0.00847 to 0.067) | 0.034 (95% interval 0.00753 to 0.062) | | Transfer Learning cohort — Source-mean comparator | 0.366 (95% interval 0.327 to 0.402) | 0.380 (95% interval 0.341 to 0.414) | −0.00664 (95% interval −0.010 to −0.00352) | −0.016 (95% interval −0.030 to −0.00526) | undefined (null) | undefined (null) | Methods & limits ## Limits ### How each result was produced ### WBCIC-SHU baselines The publisher’s processed derivative: 58 anonymous channel indices at 250 Hz in four-second epochs. Welch relative band power in 4–8, 8–13, 13–30 and 30–40 Hz per channel, as log ratios: 232 values per trial. A StandardScaler fitted on session 1 only, then a RidgeClassifier (alpha 1, no class weights). No hyperparameter search, target normalization, calibration or abstention. ### WBCIC-SHU, frozen CBraMod The same trials, resampled to 200 Hz and zero-padded to 800 samples; each trial and channel divided by its own standard deviation, a dimensionless input, not the checkpoint’s native amplitude, so upstream benchmark numbers are not reproduced. CBraMod in evaluation mode with every weight frozen; its four temporal patch embeddings averaged into 11,600 values per trial; a separate StandardScaler and RidgeClassifier fitted on each person’s session 1 only. No backbone training, LoRA or target-session calibration. ### Longitudinal RSVP 57 EEG channels, −200 to +700 ms around each event, baseline-normalized; six 100 ms averages from +100 to +700 ms, 342 values per event. Per person, first-visit blocks 1–3 fit a fixed L2 logistic regression and block 4 fits Platt calibration; later visits are scored on blocks 2–4. Group B is not used. No hyperparameter search, target-label calibration, fine-tuning or LoRA. ### Continuous cursor tracking 62 channels at 1,000 Hz, already band-passed and notch-filtered by the publisher. Five bands from 4 to 30 Hz through a per-trial FIR filter, the preceding second of band power as channel-relative log power: 310 values per row, from a trial-contained 2-second past window, every 40 ms. Per response arm, a StandardScaler and response scaling fitted on the earliest session only, then Ridge (alpha 1). No tuning, LoRA, target-session calibration or foundation-model inference. ### What the intervals cover Every interval is a pointwise 95% percentile interval from 10,000 participant (or, for cursor tracking, record) bootstrap resamples within its cohort, with paired draws for every difference. Each is conditional on the fixed trained models and predictions: refitting, preprocessing, model and protocol selection — and, for CBraMod, pretraining overlap — are not included, and there is no multiple-comparison adjustment. ### Limits that hold for all four results - The same person in every result. Each decoder is trained on that person’s own earlier labelled data and tested on their later sessions in the same dataset. - Offline only. No real-time, closed-loop, clinical or deployment claim, and in cursor tracking no claim about online control quality or intended-motion decoding. - Each result reported on its own. No single score summarises the three datasets, and none joins the eight-protocol core matrix. - Fixed protocols, frozen before scoring and not tuned on these scores; the one exception is the CBraMod arm, frozen after the spectral ridge results existed. - Target labels and metadata had been seen. Earlier WBCIC-SHU schema checks inspected label domains and counts, and RSVP event roles and target labels were inspected during quality checks. Every prediction was sealed before scoring, and no target label entered normalization, fitting, calibration, prediction or any hyperparameter, threshold, abstention or retry decision. - Publisher preprocessing as delivered. WBCIC-SHU’s channel order rests on the publisher’s procedures and the raw headers, not a direct raw-to-processed proof, and its physical amplitude is unresolved. The selection history and temporal support of its reference, filter, epoch and baseline choices are not fully established. - No calibration arm. No WBCIC-SHU arm used labelled session-3 trials, LoRA or a calibration budget, so whether a few labelled trials would help is untested here. - Cursor tracking: session, practice, speed and the historical decoder in use are confounded, and the recorded decoder labels are observational contexts, not randomized comparisons; the unknown order of decoders within a session rules out chronological-prefix and online-adaptation claims. - Audit scope. Each result passed an independent numerical audit of its saved predictions, metrics and intervals; none independently repeated the full feature extraction from the raw EEG, and for CBraMod the encoder was not rerun on all the EEG. ### Not published here Per-person and per-record values of any kind — every minimum and percentile, and every median other than the four cursor-tracking ridge medians printed beside their means; which people declined or which records did worse; WBCIC-SHU confusion matrices; the cursor-tracking results split by recorded decoder; measured compute; and any participant, record or file identifier. ### Credits, licences and consent ### WBCIC-SHU motor imagery (Yang et al. 2025) Banghua Yang, Fenqi Rong, Yunlong Xie, Du Li, Jiayang Zhang, Fu Li, Guangming Shi and Xiaorong Gao · A multi-day and high-quality EEG dataset for motor imagery brain-computer interface, Scientific Data 12, 488 (2025), doi:10.1038/s41597-025-04826-y. Data: Banghua Yang and Fenqi Rong, WBCIC-SHU Motor Imagery Dataset, Figshare+, doi:10.25452/figshare.plus.22671172.v5, CC BY 4.0. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the paper states them:** Approved by the Tsinghua University Medical Ethics Committee (approval number 20190002); carried out in line with the Declaration of Helsinki. Written informed consent was obtained from the participants after they were informed about the procedure, purpose, requirements and motor-imagery techniques. [Dataset record ↗](https://doi.org/10.25452/figshare.plus.22671172.v5) · [Paper ↗](https://doi.org/10.1038/s41597-025-04826-y) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Read from the paper’s full text ↗](https://europepmc.org/article/PMC/PMC11930978) ### Longitudinal ERP dataset, RSVP face task (Yang et al. 2025) Chen Yang, Yufeng Zhang, Hongxin Zhang, Yixuan Li, Yijun Wang and Xiaorong Gao · A longitudinal EEG dataset of event-related potential, figshare, doi:10.6084/m9.figshare.27201003.v1, CC0. Paper: Yufeng Zhang, Hongxin Zhang, Yixuan Li, Yijun Wang, Xiaorong Gao and Chen Yang, A longitudinal EEG dataset of event-related potential, Scientific Data 12, 1069 (2025), doi:10.1038/s41597-025-05378-x. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the paper states them:** Approved by the Tsinghua Institutional Review Board (No. 20230058). Each participant read and signed an informed consent form before the experiment. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.27201003.v1) · [Paper ↗](https://doi.org/10.1038/s41597-025-05378-x) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) · [Read from the paper’s full text ↗](https://europepmc.org/article/PMC/PMC12185715) ### Forenzo & He continuous-tracking EEG-BCI dataset (KiltHub) Dylan Forenzo and Bin He · EEG-BCI Dataset for “Continuous Tracking using Deep Learning-based Decoding for Non-invasive Brain-Computer Interface”, KiltHub, Carnegie Mellon University, doi:10.1184/R1/25360300.v1, CC BY 4.0. Citation the record requests: Dylan Forenzo, Hao Zhu, Jenn Shanahan, Jaehyun Lim and Bin He, Continuous tracking using deep learning-based decoding for noninvasive brain–computer interface, PNAS Nexus 3(4), pgae145 (2024), doi:10.1093/pnasnexus/pgae145. Derived analysis by BCI Report; not endorsed by the authors and not a reproduction of the paper's online results. **Ethics and consent, as the paper states them:** Approved by the Institutional Review Board at Carnegie Mellon University; no approval number is given. Each subject provided written consent to the protocol before participating. [Dataset record ↗](https://doi.org/10.1184/R1/25360300.v1) · [Paper ↗](https://doi.org/10.1093/pnasnexus/pgae145) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Read from the paper’s full text ↗](https://europepmc.org/article/PMC/PMC11060102) ### CBraMod CBraMod, the frozen encoder: its paper and the pinned checkpoint’s SHA-256 `0792cb808c14`… (revision `b9e961003214`…). Its pretraining exposure for WBCIC-SHU is not established: the model paper describes pretraining on the TUH EEG corpus and the pinned repository lists SHU as a downstream task, and neither gives a checkpoint-specific file manifest. [CBraMod paper ↗](https://proceedings.iclr.cc/paper_files/paper/2025/file/bbbd6d915cb90be21c1254a82d45cedd-Paper-Conference.pdf) · [CBraMod’s page](https://bci.report/methods/cbramod/) Derived analyses by BCI Report, aggregates only; no raw data are redistributed. Data source: [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) · schema bci-report-later-sessions-update-v1. A longer write-up of these results is in Towards AI: [Train once, reuse later: three tests of skipping BCI calibration ↗](https://pub.towardsai.net/train-once-reuse-later-three-tests-of-skipping-bci-calibration-59641e38b4ae) ## Related questions - The same person’s next session, with and without a few labelled trials from it — OpenBMI motor imagery, two classical baselines: [How much calibration data does an EEG decoder need?](https://bci.report/topics/calibration-budget/#next-session) - The next day, with three ways to update a pretrained model — run and independently replayed, and held, so no figure is published: [New people, next day: which part of a pretrained model should you update?](https://bci.report/topics/model-adaptation/#next-day) - Later sessions of one person, a longitudinal pilot listed as status only, on the same page: [Same person, later sessions](https://bci.report/topics/model-adaptation/#cross-session). - CBraMod and other frozen encoders on the core protocols, and pretrained against random initialization: [Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Does a decoder trained on an earlier session still work later?* https://bci.report/topics/later-sessions/ Figures from release [`later-sessions-update-20261008`](https://bci.report/releases/#later-sessions-update-20261008) (2026-10-08). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/later-sessions/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # How much calibration data does an EEG decoder need? Calibration budget · the same session and the next one On a 12-target wearable protocol, target-only spectral ridge improves gradually but remains low; single-band ensemble TRCA benefits sharply on wet recordings from 24 to 48 labels and only modestly on dry recordings. Across two sessions of the same people, a few labeled trials from the new session raise one motor-imagery baseline on average, though some people decline; the other baseline’s gain is not established. - [Reviewed aggregate JSON ↓](https://bci.report/data/deployment-topics.json) - Measured on: [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/); [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/) · Methods: [CCA](https://bci.report/methods/cca/); [eTRCA](https://bci.report/methods/etrca/) ## Short answer It depends on the method, and on how far the test lies from the calibration. Within one session — a 12-target wearable SSVEP task, the same **102** people tested on the same future blocks — going from 24 to 48 labeled target trials lifted single-band eTRCA from **39.8%** to **64.0%** balanced accuracy on wet recordings but only from **12.4%** to **15.1%** on dry ones, and no calibrated eTRCA point beat the training-free CCA reference (**81.1%** wet, **65.1%** dry); that eTRCA uses one fixed band, not the original method's filter bank. Across sessions — OpenBMI motor imagery, trained on each person's first session and tested on their second — **40** labeled second-session trials raised a log-covariance decoder by **+4.6 pp** on average across **51** people (interval **+2.3** to **+6.9 pp**), yet **13** of them still declined; a relative-PSD decoder gained **+1.7 pp**, but its interval includes zero, so that gain is not established, and **16** people declined. **Labeled-trial budgets for the same 102 people, tested on the same future blocks.** Participant-mean balanced accuracy on 12-target, two-second SSVEP. Dot: the estimate; line: participant bootstrap 95% interval; dashed line: chance, 1 in 12. Rust rows are wet recordings, dark rows dry. CCA uses no labels; the source-only arm is a different mechanism and is not plotted. - CCA · 0 labels — wet recording · training-free reference: 81.1% (77.2%–84.8%) - Spectral ridge · 12 labels — wet recording · target-only: 15.3% (13.7%–17.1%) - Spectral ridge · 24 labels — wet recording · target-only: 18.2% (15.9%–20.7%) - Spectral ridge · 48 labels — wet recording · target-only: 21.3% (18.3%–24.5%) - Single-band eTRCA · 24 labels — wet recording · target-only: 39.8% (34.1%–45.8%) - Single-band eTRCA · 48 labels — wet recording · target-only: 64.0% (57.1%–70.7%) - CCA · 0 labels — dry recording · training-free reference: 65.1% (60.0%–70.0%) - Spectral ridge · 12 labels — dry recording · target-only: 15.0% (13.6%–16.5%) - Spectral ridge · 24 labels — dry recording · target-only: 18.3% (16.3%–20.7%) - Spectral ridge · 48 labels — dry recording · target-only: 21.8% (19.3%–24.5%) - Single-band eTRCA · 24 labels — dry recording · target-only: 12.4% (10.4%–15.0%) - Single-band eTRCA · 48 labels — dry recording · target-only: 15.1% (12.6%–18.2%) Target-only spectral ridge ## Twelve, 24 and 48 labeled trials Each target-person model uses only that person's labeled calibration recordings. All conditions are evaluated on the same future blocks 5–10, across 102 people and 7,344 test trials. wet · 12 labels 15.3% Participant-mean balanced accuracy — 95% interval 13.7%–17.1% wet · 24 labels 18.2% Participant-mean balanced accuracy — 95% interval 15.9%–20.7% wet · 48 labels 21.3% Participant-mean balanced accuracy — 95% interval 18.3%–24.5% dry · 12 labels 15.0% Participant-mean balanced accuracy — 95% interval 13.6%–16.5% dry · 24 labels 18.3% Participant-mean balanced accuracy — 95% interval 16.3%–20.7% dry · 48 labels 21.8% Participant-mean balanced accuracy — 95% interval 19.3%–24.5% **Trial counts are not minutes.** A separate arm ## Zero target labels is source-only transfer The zero-budget spectral-ridge rows use a model trained on other participants from one source sensor. Participant-mean balanced accuracy on common future test blocks 5–10 · 0 labeled target trials. | Training source | Test on wet | Test on dry | | --- | --- | --- | | wet source — other participants only | 39.2% (35.4%–43.0%) | 35.4% (32.0%–39.0%) | | dry source — other participants only | 39.0% (35.2%–42.9%) | 37.2% (33.6%–40.9%) | eTRCA and CCA ## Single-band eTRCA responds differently The target-only eTRCA arm uses 24 or 48 calibration trials and the same future test blocks. CCA at zero labels is the analytic reference. **+24.2 pp** Wet eTRCA: 48 labels minus 24 labels ### A larger budget helped wet recordings much more The paired participant interval is +19.7 to +28.8 pp. Dry eTRCA improved by +2.7 pp, interval +1.3 to +4.2 pp. Neither calibrated point exceeds the analytic CCA reference on the same test blocks. Participant-mean balanced accuracy · native 8 channels · 2 s · 12 targets. | Method and labeled target trials | Wet | Dry | | --- | --- | --- | | CCA · 0 labels — analytic target-recording reference | 81.1% (77.2%–84.8%) | 65.1% (60.0%–70.0%) | | Single-band eTRCA · 24 labels — target-only adaptation | 39.8% (34.1%–45.8%) | 12.4% (10.4%–15.0%) | | Single-band eTRCA · 48 labels — target-only adaptation | 64.0% (57.1%–70.7%) | 15.1% (12.6%–18.2%) | Next session · OpenBMI motor imagery ## The next session: how much do a few labeled trials help? This result is kept apart from the SSVEP results above: a different data set, task and test. In OpenBMI, people imagined moving the left or the right hand in two sessions. Each person’s decoder is trained on their session 1 — the first 80 trials fit it, the last 20 calibrate its probabilities — then given the first 0, 10, 20 or 40 labeled trials of their session 2, and tested on the final 60 trials of session 2. Every budget is tested on those same trials. People evaluated 51 54 acquired, minus 1 engineering exclusion and 2 held for input quality. Held-out trials 3,060 60 per person, the final trials of session 2. Prediction jobs 408 Every person × two baselines × four budgets; none failed. Participant-mean balanced accuracy on the same final trials of session 2, 51 people, with whole-participant bootstrap 95% intervals. Chance is 50.0%. Accuracy and macro F1 for every cell are in the download. | Labeled session-2 trials | Log-covariance + shrinkage LDA | Relative PSD + standardized ridge | | --- | --- | --- | | None · session 1 only | 67.9% (95% interval 63.6%–72.4%) | 57.3% (95% interval 54.4%–60.5%) | | 10 trials | 69.1% (95% interval 64.8%–73.5%) | 57.4% (95% interval 54.5%–60.5%) | | 20 trials | 69.2% (95% interval 65.3%–73.4%) | 57.9% (95% interval 55.2%–60.8%) | | 40 trials | 72.4% (95% interval 68.9%–76.2%) | 59.0% (95% interval 56.0%–62.1%) | **+4.6 pp** Log-covariance LDA: 40 labeled session-2 trials minus none, the same 51 people ### Better on average — and still worse for some people Paired interval +2.3 to +6.9 pp, above zero. Yet 13 of 51 people scored lower with the trials than without them, 6 of them by 5 points or more. For the relative-PSD baseline the mean change was +1.7 pp (−0.1 to +3.5 pp): the interval includes zero, so that gain is not established; 16 of 51 people declined, 6 by 5 points or more. Change in participant-mean balanced accuracy against no session-2 trials, the same people and test trials, with paired whole-participant bootstrap 95% intervals — and, beside each mean, how many people declined. | Comparison | Log-covariance + shrinkage LDA | Relative PSD + standardized ridge | | --- | --- | --- | | 10 trials minus none | +1.2 pp (95% interval −0.6 to +3.3 pp) (21 of 51 declined · 8 by 5 points or more) (The interval includes zero: not established.) | +0.1 pp (95% interval −1.1 to +1.2 pp) (25 of 51 declined · 6 by 5 points or more) (The interval includes zero: not established.) | | 20 trials minus none | +1.4 pp (95% interval −0.3 to +3.2 pp) (24 of 51 declined · 6 by 5 points or more) (The interval includes zero: not established.) | +0.6 pp (95% interval −0.9 to +2.0 pp) (19 of 51 declined · 10 by 5 points or more) (The interval includes zero: not established.) | | 40 trials minus none | +4.6 pp (95% interval +2.3 to +6.9 pp) (13 of 51 declined · 6 by 5 points or more) | +1.7 pp (95% interval −0.1 to +3.5 pp) (16 of 51 declined · 6 by 5 points or more) (The interval includes zero: not established.) | An expanded cohort under the same fixed method, not an independent replication: the original 40 people, their results unchanged, plus all 11 people who became eligible later. An earlier snapshot of the original people was prepared but never published here. ### Limits of the next-session result ### One dataset, offline Online closed-loop control, clinical efficacy, consumer headsets and reduced montages are outside this test. ### Same person Every model is fitted to the person it is tested on. ### Two fixed CPU baselines No parameter search, score-driven retry or trial dropping, and no foundation model or fine-tuning (LoRA/PEFT). They are a reference, not a ceiling. ### Noisy per person Each person’s change rests on 60 test trials, which the four budgets reuse, so one person’s decline can be noise. ### Units unresolved The recordings’ physical amplitude unit is not resolved, so only dimensionless inputs are used: each trial filtered 8–30 Hz, covariances divided by their trace, band powers as shares. ### Pointwise intervals Whole-participant bootstrap, 10,000 draws shared by every baseline and budget, not adjusted for multiple comparisons. ### OpenBMI motor imagery (Lee et al. 2019) Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. The paper’s Ethical Approval section states approval by the Korea University Institutional Review Board (1040548-KUIRB-16-159-A-2) and written informed consent from all participants before the experiments. Only cohort means, paired changes and how many people declined appear here — never who. [GigaDB dataset record ↗](https://doi.org/10.5524/100542) · [GigaScience paper ↗](https://doi.org/10.1093/gigascience/giz002) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) Data source: [large-source-update.json](https://bci.report/data/large-source-update.json) · schema bci-report-large-source-update-v1. ## LaBraM adaptation has its own page The LaBraM results on new people, same task, zero labels from the test person — head only, last block or LoRA — moved to their own page: [New people, next day: which part of a pretrained model should you update?](https://bci.report/topics/model-adaptation/#adaptation) ## The next-day experiment has moved to the model-adaptation page The next-day status — run, independently replayed and held, with no figure published — is on the same page: [New people, next day: which part of a pretrained model should you update?](https://bci.report/topics/model-adaptation/#next-day) Methods & limits ## Limits of the SSVEP budget curves ### Prespecified single band eTRCA uses the fixed 6–80 Hz prepared band, not the filter bank of the original method (five sub-bands in the paper and in the reference code’s tutorials), so low scores do not establish an optimized TRCA ceiling. ### Descriptive intervals Bootstrap intervals summarize participants under this protocol, without retraining uncertainty. ### An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces. Sensors, 21(4), 1256. https://doi.org/10.3390/s21041256 Official Tsinghua author mirror snapshot acquired 2026-09-20 · CC BY 4.0. Aggregate results only. [Dataset record ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) · [Sensors paper ↗](https://doi.org/10.3390/s21041256) Data source: [reviewed aggregate JSON](https://bci.report/data/deployment-topics.json) · schema bci-report-public-deployment-topics-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *How much calibration data does an EEG decoder need?* https://bci.report/topics/calibration-budget/ Figures from releases [`large-source-update-20261003`](https://bci.report/releases/#large-source-update-20261003) (2026-10-03), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/calibration-budget/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # New people, next day: which part of a pretrained model should you update? Model adaptation · new people; next day held Which part of a pretrained encoder to update for people it has not seen, on a task it is adapted to: the head only, the last block, or a low-rank adapter, each from the same checkpoint and five-epoch recipe. The next-day version has run; its source is held. - [Reviewed aggregate JSON ↓](https://bci.report/data/adaptation-update.json) - Measured on: [EEGMAT](https://bci.report/datasets/eegmat/) · Methods: [LaBraM](https://bci.report/methods/labram/) ## Short answer On new people, same task, zero labels from the test person — LaBraM Base on mental arithmetic versus rest, **36** people in five participant-disjoint folds — updating the last block with the head reached **65.7%** balanced accuracy and rank-4 LoRA **64.1%**, against **56.6%** for a short head-only fit; the paired interval for LoRA minus last block crosses zero. On the same people and folds, the core matrix's frozen LaBraM with a ridge head scored **64.6%**. The next-day experiment ran, but its source is held, so its figures are unpublished. New people, same task, zero labels from the test person ## Adapting a foundation model: head only, last block or LoRA? LaBraM Base on EEGMAT, mental arithmetic versus rest: 36 people in five participant-disjoint folds. The three update rules start from the same head, see the same batches in the same order and train for the same five epochs; three seeds are averaged within each person before the cohort mean. Head only · encoder frozen 56.6% Macro F1 54.6% · 402 trainable parameters — 95% interval 54.2%–59.0% Last block + head 65.7% Macro F1 62.3% · 482,882 trainable parameters — 95% interval 61.6%–69.7% LoRA rank 4 + head 64.1% Macro F1 60.5% · 38,802 trainable parameters — 95% interval 60.2%–67.8% For scale: the core matrix's frozen LaBraM, on the same people and folds but read out by a closed-form ridge head on standardized features, scored 64.6% (core matrix release, [experiments.json](https://bci.report/data/experiments.json)). The head-only arm here is a short five-epoch head on unscaled features — not the best a frozen encoder can do — and the two are not a paired comparison. **+7.5 pp** LoRA minus head only, same people ### LoRA beat the short head-only fit; it did not beat the last block 30 of 36 people improved and 6 got worse; paired interval +4.8 to +10.1 pp. Updating the last block instead gained +9.1 pp (+6.0 to +12.2 pp). LoRA minus last block was −1.6 pp, interval −4.4 to +1.0 pp: it crosses zero, so neither is shown to be better. 17 people did better with LoRA, 19 with the last block. Change in participant-mean score for the same 36 people under both rules · descriptive person-bootstrap 95% interval · people helped, harmed and unchanged. | Comparison | Balanced accuracy | Macro F1 | | --- | --- | --- | | LoRA − head only | +7.5 pp (+4.8 to +10.1 pp) (30 helped · 6 harmed · 0 unchanged) | +5.9 pp (+2.7 to +9.0 pp) (22 helped · 14 harmed · 0 unchanged) | | Last block − head only | +9.1 pp (+6.0 to +12.2 pp) (30 helped · 5 harmed · 1 unchanged) | +7.7 pp (+4.4 to +11.2 pp) (27 helped · 9 harmed · 0 unchanged) | | LoRA − last block | −1.6 pp (−4.4 to +1.0 pp) (17 helped · 19 harmed · 0 unchanged) | −1.9 pp (−5.0 to +1.1 pp) (15 helped · 21 harmed · 0 unchanged) | On macro F1 the gains are smaller, and LoRA helps fewer people over head only than it does on balanced accuracy. ### What each rule costs Training time is the sum over 15 fits (5 folds × 3 seeds) on one Apple-silicon Mac, PyTorch MPS backend, excluding loading and evaluation. | Update rule | Trainable parameters | Training time, 15 fits | Balanced accuracy, seed by seed | | --- | --- | --- | --- | | Head only · encoder frozen | 402 | 51.5 s | (56.9% · 56.4% · 56.5%) | | Last block + head | 482,882 | 65.3 s | (65.7% · 65.7% · 65.7%) | | LoRA rank 4 + head | 38,802 | 128.2 s | (64.4% · 65.1% · 62.7%) | LoRA trains far fewer weights than the last block but took longer here: its adapters sit in all twelve blocks, so gradients flow back through most of the encoder, and this implementation is an effective-weight parametrization, not an optimized low-rank kernel. Promised on the roadmap and not delivered by this batch: peak memory (only lower-bound samples were recorded, so none is published) and a comparison at a fixed computation budget (every rule trained for the same five epochs, not the same compute). Nine further foundation encoders were also adapted on these same people and folds with two of the rules above — a head trained on the frozen encoder, and rank-4 LoRA — and the same five-epoch recipe, with LoRA budgets that differ by model: [their table is on “Does pretraining help?”](https://bci.report/topics/does-pretraining-help/#v9-adaptation). From the saved test scores of the head-only and LoRA arms above: how good their probabilities are — log loss and calibration error on the same people — and when either should decline a decision. Both are on [“When should a decoder decline to decide?”](https://bci.report/topics/when-not-to-act/#rd-lora). Next day · status only ## Same person, the next day Does a model fitted on one day still work on a later one, and which part should be updated when it does not? Two sources were prepared for it; for different reasons, neither can be scored in public. ### Two recording days: run and replayed, held BNCI2015-001: 12 people recorded on two days. Each person's model is trained on day A, calibrated with the first 0, 10, 20 or 40 labeled trials of day B under the same three update rules, and tested on the same later day-B trials. Unlike the result above, this design uses the test person's own labels: their day-A trials to train and, except at the zero budget, their first day-B trials to calibrate. It was run and passed an independent replay. The source's editorial hold stands — its catalogue licence is CC BY-NC-ND and its description names no ethics approval — so no figure from it is published. ### Same person, later sessions: status only Stieger longitudinal BCI: one person's 11 sessions were acquired, a subset of the release. The loader and a chronological split work on it: earlier sessions train, later sessions test. With one person, every score would be that person's, so none is published; the source is listed as status only in the [release log](https://bci.report/releases/#context-update-20260927). A related question now has a measured answer, for two classical baselines on other data: the same person’s next session, with and without a few labeled trials from it. It is on [How much calibration data does an EEG decoder need?](https://bci.report/topics/calibration-budget/#next-session). Methods & limits ## Scope of this result ### One untuned recipe Five epochs, AdamW at learning rate 1e-4, batch 32, with no search, validation split, early stopping or checkpoint selection. Another learning rate or training budget could reorder the rules. ### New people, same task The folds hold out people on a task the model is adapted to. Nothing here measures a new recording day, a new headset or a new dataset. ### A known dataset EEGMAT is public. It is not in LaBraM's published pretraining list, checked on 4 October 2026 (source below): what the authors' list shows, not proof that the recordings were never seen. ### Descriptive intervals Person bootstrap after averaging the three seeds within each person. It ignores the dependence that shared cross-validation models create. LaBraM’s pretraining list, checked 2026-10-04: [source ↗](https://arxiv.org/abs/2405.18765) ### Adapter size The LaBraM base encoder has 5,819,936 parameters. Rank-4 updates to its twelve fused QKV weights add 38,400 trainable parameters (0.66%); with the 402-parameter head that is the 38,802 above. The adapter tensors take 150 KiB in float32 before metadata. **Engineering checks passed before any real-EEG run:** exact adapter save and reload; exact merged-weight equivalence; finite, nonzero adapter gradients in every block; frozen encoder weights unchanged after updates; identical output at zero adapter initialization. These counts describe the adapter’s size only. ### Related work Other groups already benchmark EEG adapters: [OpenEEGBench](https://github.com/braindecode/OpenEEGBench) covers parameter-efficient fine-tuning of EEG foundation models; a [systematic study of test-time adaptation](https://proceedings.mlr.press/v340/lee26a.html) reports inconsistent gains under distribution shift; a recent [preprint](https://arxiv.org/abs/2608.24727v1) compares self-supervised adaptation under fixed computation budgets. ### EEGMAT · LaBraM adaptation, three matched update rules Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. Open Data Commons Attribution License 1.0 · aggregate results only; the adaptation reuses the 2026-09-20 review of these recordings. [Dataset record ↗](https://physionet.org/content/eegmat/1.0.0/) ### Stieger longitudinal BCI · one-person pilot James R. Stieger, Stephen A. Engel and Bin He · Continuous sensorimotor rhythm based brain computer interface learning in a large population, Scientific Data 8, 98 (2021); data at doi:10.6084/m9.figshare.13123148.v1. CC-BY-4.0 · status only. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.13123148.v1) Data source: [adaptation-update.json](https://bci.report/data/adaptation-update.json) · schema bci-report-adaptation-update-v1. Data source: [evidence-update.json](https://bci.report/data/evidence-update.json) · the adapter counts and checks, from the released engineering check. A longer write-up of the EEGMAT adaptation results is on Hugging Face: [Choosing how to adapt an EEG foundation model ↗](https://huggingface.co/blog/Twu31/choosing-how-to-adapt-an-eeg-foundation-model) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *New people, next day: which part of a pretrained model should you update?* https://bci.report/topics/model-adaptation/ Figures from releases [`adaptation-update-20261001`](https://bci.report/releases/#adaptation-update-20261001) (2026-10-01), [`evidence-update-20260922`](https://bci.report/releases/#evidence-update-20260922) (2026-09-22), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/model-adaptation/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Does pretraining help EEG foundation models like LaBraM and CBraMod? Representation controls · matched frozen encoders Sometimes. CBraMod and EEGPT gain under both readout settings on both tasks; LaBraM gains on mental workload and does not on MI/rest. Twelve further foundation models, LaBraM at Scale’s Quark ladder among them, are compared in the same frozen recipe. Five of their checkpoints also have random-initialization controls. - [Reviewed aggregate JSON ↓](https://bci.report/data/deployment-topics.json) - Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Methods: [EEGNet](https://bci.report/methods/eegnet/); [LaBraM](https://bci.report/methods/labram/); [CBraMod](https://bci.report/methods/cbramod/); [CCA](https://bci.report/methods/cca/); [REVE](https://bci.report/methods/reve/); [LUNA](https://bci.report/methods/luna/); [BrainOmni](https://bci.report/methods/brainomni/); [CodeBrain](https://bci.report/methods/codebrain/); [EEGMamba](https://bci.report/methods/eegmamba/); [ST-EEGFormer](https://bci.report/methods/st-eegformer/); [eeg-fm-masking](https://bci.report/methods/eeg-fm-masking/); [ERP-FM](https://bci.report/methods/erp-fm/); [SingLEM](https://bci.report/methods/singlem/); [ZUNA](https://bci.report/methods/zuna/); [EEGPT](https://bci.report/methods/eegpt/) ## Short answer Sometimes, depending on the encoder and the task: with frozen encoders and the fixed readout, pretrained CBraMod beat the mean of its three random initializations by **+6.39 pp** of balanced accuracy on MI/rest (**10** people) and by **+5.90 pp** on mental workload (**36** people), pretrained EEGPT beat its own by **+4.25 pp** and **+6.45 pp** on the same two tasks, and pretrained LaBraM beat its own by **+8.12 pp** on mental workload, each with a paired 95% interval above zero. On MI/rest, LaBraM's difference was **−2.42 pp** (interval crosses zero), and **−3.56 pp** under the train-selected readout (interval wholly below zero). Neither benchmark is in LaBraM's, CBraMod's or EEGPT's published pretraining list (checked 2026-10-04, EEGPT's 2026-10-12). The evidence is for these two tasks and frozen encoders. Twelve further foundation models — twenty-three encoder checkpoints, LaBraM at Scale’s Quark ladder and REVE Small’s early checkpoint among them — were run in the same frozen recipe; five of those checkpoints, of four models (REVE Large, ST-EEGFormer Base and Large, SingLEM and EEGMamba), also have a random-initialization control on motor imagery and rest, mental workload and sleep staging (below), and the others have none. Under the fixed readout, **10** of their **15** paired intervals lie wholly above zero, **1** wholly below (SingLEM on sleep staging) and **4** include zero. On sleep staging (**20** people) REVE Large, ST-EEGFormer Base and ST-EEGFormer Large (**76.6%**, **77.3%**, **79.6%**) lie above every published row, spectral ridge at **72.2%** included, as does EEGPT, whose interval clears spectral ridge's by only **0.10 pp**; on BETA none lies above training-free CCA (**63.1%**); on EEGMAT the frozen readouts and LoRA order the encoders differently. **Pretrained minus random: MI/rest, 10 people; mental workload, 36 people.** Difference in participant-mean balanced accuracy, in percentage points (pp): pretrained encoder minus the mean of three random initializations, same frozen architecture and splits. Dot: the difference; line: paired participant bootstrap 95% interval, printed under each label; dashed line: zero, no difference. Dark rows use the fixed readout, rust rows the train-selected readout. - LaBraM · MI / rest — Fixed readout · −5.94 to +1.03 pp: −2.42 pp - LaBraM · MI / rest — Train-selected readout · −7.00 to −0.19 pp: −3.56 pp - CBraMod · MI / rest — Fixed readout · +3.06 to +9.50 pp: +6.39 pp - CBraMod · MI / rest — Train-selected readout · +4.53 to +11.11 pp: +7.89 pp - EEGPT · MI / rest — Fixed readout · +0.75 to +7.44 pp: +4.25 pp - EEGPT · MI / rest — Train-selected readout · +0.72 to +7.08 pp: +3.92 pp - LaBraM · Mental workload — Fixed readout · +4.97 to +11.40 pp: +8.12 pp - LaBraM · Mental workload — Train-selected readout · +4.74 to +11.42 pp: +7.99 pp - CBraMod · Mental workload — Fixed readout · +2.31 to +9.57 pp: +5.90 pp - CBraMod · Mental workload — Train-selected readout · +3.12 to +10.43 pp: +6.73 pp - EEGPT · Mental workload — Fixed readout · +2.61 to +10.59 pp: +6.45 pp - EEGPT · Mental workload — Train-selected readout · +4.40 to +10.40 pp: +7.38 pp Paired attribution contrasts ## Pretrained minus random, same readout Each constructor-random architecture is averaged over three initialization seeds. Pretrained and random encoders share the same splits, pooling and ridge readout protocol. Difference in participant-mean balanced accuracy. Intervals are paired participant bootstraps, conditional on three random initializations. | Task / model | Fixed alpha 100 | Train-selected alpha | | --- | --- | --- | | LaBraM — MI / rest · n=10 | −2.42 pp (−5.94 to +1.03 pp) | −3.56 pp (−7.00 to −0.19 pp) | | CBraMod — MI / rest · n=10 | +6.39 pp (+3.06 to +9.50 pp) | +7.89 pp (+4.53 to +11.11 pp) | | EEGPT — MI / rest · n=10 | +4.25 pp (+0.75 to +7.44 pp) | +3.92 pp (+0.72 to +7.08 pp) | | LaBraM — Mental workload · n=36 | +8.12 pp (+4.97 to +11.40 pp) | +7.99 pp (+4.74 to +11.42 pp) | | CBraMod — Mental workload · n=36 | +5.90 pp (+2.31 to +9.57 pp) | +6.73 pp (+3.12 to +10.43 pp) | | EEGPT — Mental workload · n=36 | +6.45 pp (+2.61 to +10.59 pp) | +7.38 pp (+4.40 to +10.40 pp) | Measured scores ## Scores under both readout settings Head selection uses only inner training participants. Fixed head · alpha 100 · participant-mean balanced accuracy. | Task / model | Pretrained encoder | Random encoder mean | Paired difference | | --- | --- | --- | --- | | LaBraM — MI / rest | 53.5% | 55.9% — 3 initializations | −2.42 pp (−5.94 to +1.03 pp) | | CBraMod — MI / rest | 62.9% | 56.5% — 3 initializations | +6.39 pp (+3.06 to +9.50 pp) | | EEGPT — MI / rest | 58.3% | 54.1% — 3 initializations | +4.25 pp (+0.75 to +7.44 pp) | | LaBraM — Mental workload | 64.6% | 56.5% — 3 initializations | +8.12 pp (+4.97 to +11.40 pp) | | CBraMod — Mental workload | 62.3% | 56.4% — 3 initializations | +5.90 pp (+2.31 to +9.57 pp) | | EEGPT — Mental workload | 65.7% | 59.2% — 3 initializations | +6.45 pp (+2.61 to +10.59 pp) | Train-selected head · alpha 100 / 10,000 / 1 · participant-mean balanced accuracy. | Task / model | Pretrained encoder | Random encoder mean | Paired difference | | --- | --- | --- | --- | | LaBraM — MI / rest | 52.9% | 56.5% — 3 initializations | −3.56 pp (−7.00 to −0.19 pp) | | CBraMod — MI / rest | 62.7% | 54.9% — 3 initializations | +7.89 pp (+4.53 to +11.11 pp) | | EEGPT — MI / rest | 57.0% | 53.1% — 3 initializations | +3.92 pp (+0.72 to +7.08 pp) | | LaBraM — Mental workload | 64.4% | 56.5% — 3 initializations | +7.99 pp (+4.74 to +11.42 pp) | | CBraMod — Mental workload | 68.4% | 61.7% — 3 initializations | +6.73 pp (+3.12 to +10.43 pp) | | EEGPT — Mental workload | 66.0% | 58.6% — 3 initializations | +7.38 pp (+4.40 to +10.40 pp) | EEGPT’s rows: its pretrained rows under the fixed head are its core-matrix cells on these two protocols, the same predictions, so the notes on those cells hold here. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ds003810 is recorded at 125 Hz; its windows reach the encoder upsampled to 256 Hz, so they carry nothing above 62.5 Hz. Fixed classical control 52.3% Log-covariance ridge · MI / rest — 95% interval 48.5%–56.4% Fixed classical control 61.9% Log-covariance ridge · Mental workload — 95% interval 56.2%–67.6% Paired attribution contrasts · further encoders ## Pretrained minus random, further encoders REVE Large, ST-EEGFormer Base and Large, SingLEM and EEGMamba, each against the mean of 3 constructor-random initializations of its own architecture, on motor imagery and rest (10 people), mental workload (36) and sleep staging (20): the same windows, folds, pooling and ridge head as their published rows, so only the weights differ. Under the fixed readout, 10 of the 15 paired intervals lie wholly above zero, 1 wholly below (SingLEM on sleep staging) and 4 include zero. **Pretrained minus random, by task and encoder.** Difference in participant-mean balanced accuracy, in percentage points (pp): pretrained encoder minus the mean of its random initializations, same frozen architecture, people and folds. Dot: the difference; line: paired participant bootstrap 95% interval, printed under each label; dashed line: zero. Dark rows use the fixed readout, rust rows the train-selected readout. This axis is wider than the one above. - REVE Large · MI / rest — Fixed readout · −6.06 to +8.72 pp: +2.11 pp - REVE Large · MI / rest — Train-selected readout · −8.11 to +3.89 pp: −1.75 pp - ST-EEGFormer Base · MI / rest — Fixed readout · −1.78 to +7.11 pp: +2.25 pp - ST-EEGFormer Base · MI / rest — Train-selected readout · −0.11 to +7.56 pp: +3.67 pp - ST-EEGFormer Large · MI / rest — Fixed readout · +1.97 to +13.08 pp: +7.42 pp - ST-EEGFormer Large · MI / rest — Train-selected readout · −0.31 to +10.75 pp: +5.11 pp - SingLEM · MI / rest — Fixed readout · +11.14 to +20.20 pp: +15.89 pp - SingLEM · MI / rest — Train-selected readout · +11.72 to +21.17 pp: +16.58 pp - EEGMamba · MI / rest — Fixed readout · +6.47 to +9.25 pp: +7.86 pp - EEGMamba · MI / rest — Train-selected readout · +2.69 to +9.97 pp: +6.33 pp - REVE Large · Mental workload — Fixed readout · +9.95 to +21.44 pp: +15.77 pp - REVE Large · Mental workload — Train-selected readout · +10.26 to +21.02 pp: +15.59 pp - ST-EEGFormer Base · Mental workload — Fixed readout · +0.85 to +10.65 pp: +5.62 pp - ST-EEGFormer Base · Mental workload — Train-selected readout · +1.37 to +9.52 pp: +5.29 pp - ST-EEGFormer Large · Mental workload — Fixed readout · −0.65 to +9.68 pp: +4.40 pp - ST-EEGFormer Large · Mental workload — Train-selected readout · +2.44 to +12.73 pp: +7.53 pp - SingLEM · Mental workload — Fixed readout · −2.89 to +2.16 pp: −0.35 pp - SingLEM · Mental workload — Train-selected readout · −1.28 to +3.55 pp: +1.13 pp - EEGMamba · Mental workload — Fixed readout · +1.74 to +6.40 pp: +4.06 pp - EEGMamba · Mental workload — Train-selected readout · 0.00 to +4.89 pp: +2.45 pp - REVE Large · Sleep staging — Fixed readout · +30.86 to +35.04 pp: +32.79 pp - REVE Large · Sleep staging — Train-selected readout · +29.73 to +34.17 pp: +31.80 pp - ST-EEGFormer Base · Sleep staging — Fixed readout · +7.58 to +10.39 pp: +8.99 pp - ST-EEGFormer Base · Sleep staging — Train-selected readout · +7.99 to +11.01 pp: +9.53 pp - ST-EEGFormer Large · Sleep staging — Fixed readout · +9.12 to +12.60 pp: +10.84 pp - ST-EEGFormer Large · Sleep staging — Train-selected readout · +8.87 to +11.96 pp: +10.44 pp - SingLEM · Sleep staging — Fixed readout · −11.13 to −7.40 pp: −9.30 pp - SingLEM · Sleep staging — Train-selected readout · −12.61 to −9.61 pp: −11.06 pp - EEGMamba · Sleep staging — Fixed readout · +5.08 to +8.11 pp: +6.60 pp - EEGMamba · Sleep staging — Train-selected readout · +5.19 to +8.17 pp: +6.68 pp Fixed readout · The published rows’ head: ridge, alpha 100, on features standardized on the training folds, class-balanced. Participant-mean balanced accuracy; the difference and its paired 95% interval in percentage points. | Task / encoder | Pretrained encoder | Random encoder mean | Paired difference | Reading | | --- | --- | --- | --- | --- | | REVE Large — MI / rest · n=10 | 70.2% | 68.1% | +2.11 pp (−6.06 to +8.72 pp) | no difference shown | | ST-EEGFormer Base — MI / rest · n=10 | 58.1% | 55.8% | +2.25 pp (−1.78 to +7.11 pp) | no difference shown | | ST-EEGFormer Large — MI / rest · n=10 | 62.4% | 55.0% | +7.42 pp (+1.97 to +13.08 pp) | pretrained higher | | SingLEM — MI / rest · n=10 | 72.0% | 56.1% | +15.89 pp (+11.14 to +20.20 pp) | pretrained higher | | EEGMamba — MI / rest · n=10 | 67.3% | 59.4% | +7.86 pp (+6.47 to +9.25 pp) | pretrained higher | | REVE Large — Mental workload · n=36 | 69.0% | 53.3% | +15.77 pp (+9.95 to +21.44 pp) | pretrained higher | | ST-EEGFormer Base — Mental workload · n=36 | 64.5% | 58.9% | +5.62 pp (+0.85 to +10.65 pp) | pretrained higher | | ST-EEGFormer Large — Mental workload · n=36 | 63.4% | 59.0% | +4.40 pp (−0.65 to +9.68 pp) | no difference shown | | SingLEM — Mental workload · n=36 | 51.5% | 51.9% | −0.35 pp (−2.89 to +2.16 pp) | no difference shown — Its published pretrained interval includes chance. | | EEGMamba — Mental workload · n=36 | 60.3% | 56.2% | +4.06 pp (+1.74 to +6.40 pp) | pretrained higher | | REVE Large — Sleep staging · n=20 | 76.6% | 43.8% | +32.79 pp (+30.86 to +35.04 pp) | pretrained higher | | ST-EEGFormer Base — Sleep staging · n=20 | 77.3% | 68.3% | +8.99 pp (+7.58 to +10.39 pp) | pretrained higher | | ST-EEGFormer Large — Sleep staging · n=20 | 79.6% | 68.8% | +10.84 pp (+9.12 to +12.60 pp) | pretrained higher | | SingLEM — Sleep staging · n=20 | 35.3% | 44.6% | −9.30 pp (−11.13 to −7.40 pp) | random initializations higher | | EEGMamba — Sleep staging · n=20 | 64.1% | 57.5% | +6.60 pp (+5.08 to +8.11 pp) | pretrained higher | Train-selected readout · Alpha chosen among 100, 10,000 and 1 on inner-validation people drawn from the outer-training people only, then refit on all outer-training people. Participant-mean balanced accuracy; the difference and its paired 95% interval in percentage points. | Task / encoder | Pretrained encoder | Random encoder mean | Paired difference | Reading | | --- | --- | --- | --- | --- | | REVE Large — MI / rest · n=10 | 65.6% | 67.3% | −1.75 pp (−8.11 to +3.89 pp) | no difference shown | | ST-EEGFormer Base — MI / rest · n=10 | 59.1% | 55.4% | +3.67 pp (−0.11 to +7.56 pp) | no difference shown | | ST-EEGFormer Large — MI / rest · n=10 | 60.8% | 55.7% | +5.11 pp (−0.31 to +10.75 pp) | no difference shown | | SingLEM — MI / rest · n=10 | 72.0% | 55.4% | +16.58 pp (+11.72 to +21.17 pp) | pretrained higher | | EEGMamba — MI / rest · n=10 | 64.8% | 58.4% | +6.33 pp (+2.69 to +9.97 pp) | pretrained higher | | REVE Large — Mental workload · n=36 | 67.7% | 52.1% | +15.59 pp (+10.26 to +21.02 pp) | pretrained higher | | ST-EEGFormer Base — Mental workload · n=36 | 64.3% | 59.0% | +5.29 pp (+1.37 to +9.52 pp) | pretrained higher | | ST-EEGFormer Large — Mental workload · n=36 | 65.6% | 58.1% | +7.53 pp (+2.44 to +12.73 pp) | pretrained higher | | SingLEM — Mental workload · n=36 | 52.7% | 51.6% | +1.13 pp (−1.28 to +3.55 pp) | no difference shown | | EEGMamba — Mental workload · n=36 | 61.9% | 59.5% | +2.45 pp (0.00 to +4.89 pp) | no difference shown — The lower bound is exactly zero in exact arithmetic, so the interval includes zero. | | REVE Large — Sleep staging · n=20 | 75.6% | 43.8% | +31.80 pp (+29.73 to +34.17 pp) | pretrained higher | | ST-EEGFormer Base — Sleep staging · n=20 | 77.3% | 67.8% | +9.53 pp (+7.99 to +11.01 pp) | pretrained higher | | ST-EEGFormer Large — Sleep staging · n=20 | 79.1% | 68.7% | +10.44 pp (+8.87 to +11.96 pp) | pretrained higher | | SingLEM — Sleep staging · n=20 | 33.8% | 44.9% | −11.06 pp (−12.61 to −9.61 pp) | random initializations higher | | EEGMamba — Sleep staging · n=20 | 64.1% | 57.4% | +6.68 pp (+5.19 to +8.17 pp) | pretrained higher | Train-selected readout · Sleep staging: A five-class extension of the published binary train-selected readout, with no published reproduction. Row notes, as on the published rows: REVE Large: Input removes the per-segment per-channel mean, as in the published row. ST-EEGFormer Large: The authors’ downstream class uses a 1-based temporal index and Large may have been pretrained with a 0-based one, so its temporal sinusoid may sit one position off; this concerns the pretrained arm’s input handling. SingLEM: Each channel is encoded independently, and the tokenizer leaves the last 0.25 s of each 2 s segment unseen, as in the published row. ### Limits of these controls - Each difference is for frozen encoders read by the published ridge head (alpha 100 on features standardized on the training folds, class-balanced); fine-tuning and adaptation are separate questions. - The random arm is each upstream constructor’s own initialization in the pinned environment (for example, REVE’s linear weights drawn from N(0, 0.02) and ST-EEGFormer’s channel-embedding projection starting at zero); three initializations with the published controls’ seeds, mean only, and the intervals are paired participant bootstraps conditional on those three. - Both arms of an encoder share its input view, pooling and feature width. Widths differ across encoders (EEGMamba keeps channels × 200 dimensions; SingLEM concatenates per-channel tokens), and random features can score well under a wide linear readout, so each difference is read within its encoder. - The train-selected readout changes the reading of three cells: ST-EEGFormer Large on MI/rest and on mental workload, and EEGMamba on mental workload. - Cohorts of 10 people on MI/rest, 36 on mental workload and 20 on sleep staging; the intervals are descriptive participant bootstraps that ignore cross-validation dependence, and with 15 intervals without multiplicity correction about one in twenty can exclude zero by chance. - Same task and recording setup with held-out people; cross-dataset, cross-day and deployment questions lie outside these results. ### How these were run - The published frozen-probe recipe with only the weights changed: the published 250 Hz microvolt windows (2 s; a 30 s sleep epoch is the mean of the pooled features of 15 two-second segments), recorded channels only, each encoder’s published input view and pooling, strict fp32, and the published participant-disjoint folds, five per task. - Each encoder’s own upstream constructor, built on CPU after fixed seeds (the published controls’ seeds); every parameter and buffer of the encoder is replaced, and the input view, channel tables, electrode positions, pooling, head and folds stay as published. - 10,000 paired participant draws (seed 20260920), 2.5 and 97.5 percentiles; each draw keeps a person’s pretrained and random values together. - Every arm ran on Local Ubuntu / CUDA with deterministic settings; the pretrained and random arms of a cell share the device and the session. - Each pretrained arm was re-extracted and matched its published cell before its random arms ran: identical predictions in all 15 cells, features bitwise equal in 12 and within 2.6e-6 for EEGMamba (CUDA kernel arithmetic). - SingLEM is its downstream-excluded checkpoint (singlem_downstream_excluded.pt), the checkpoint of the published row; ST-EEGFormer Base and Large are the official release files checkpoint-288.pth and large_weights_only_196.pth. - The published head reproduced all 15 published cells from their saved features; every random state matched its digest, frozen before the run; the train-selected head reproduced the 4 published LaBraM and CBraMod train-selected values. - An export check rebuilt one of REVE Large’s three random initializations on CPU, re-extracted 64 of the 1,200 motor-imagery windows and matched the sealed features. Exposure: None of the three datasets is in these four models’ published pretraining lists (checked 2026-10-04; EEGMamba’s list read from its official code). Terms: REVE Large under the REVE Responsible Use License v1.0 (reve-large @ 317531c7): aggregate scientific results only. ST-EEGFormer, SingLEM and EEGMamba under MIT. Downloaded and random weights stay local. Data source: [random-init-controls-update.json](https://bci.report/data/random-init-controls-update.json) · schema bci-report-random-init-controls-update-v1. Training-seed sensitivity ## EEGNet across three training seeds Five existing EEGNet protocols were rerun with three fixed MPS seeds. The range is a small-sample check of training sensitivity, not a population interval; the published point estimates stand. Balanced accuracy across three training seeds. Sample SD and full range are percentage points. | Protocol | Three runs | Mean | Sample SD | Full range | | --- | --- | --- | --- | --- | | Semantic ERP — n=30 · 3,600 trials | 60.50% / 60.42% / 60.33% | 60.42% | 0.08 pp | 0.17 pp — 60.33%–60.50% | | P300 — n=21 · 2,520 trials | 55.79% / 50.36% / 52.06% | 52.74% | 2.78 pp | 5.44 pp — 50.36%–55.79% | | Scalp sleep staging — n=20 · 3,000 trials | 57.33% / 54.43% / 51.60% | 54.46% | 2.87 pp | 5.73 pp — 51.60%–57.33% | | BETA · 4 selected channels — n=70 · 11,200 trials | 44.77% / 44.42% / 45.03% | 44.74% | 0.30 pp | 0.61 pp — 44.42%–45.03% | | BETA · 8 selected channels — n=70 · 11,200 trials | 56.54% / 56.73% / 56.99% | 56.75% | 0.23 pp | 0.46 pp — 56.54%–56.99% | Dataset authors, versions, licenses and source records for all four seed-sensitivity datasets are listed under [Data use & privacy → Core and deployment sources](https://bci.report/data-use/#sources). Frozen-probe comparison ## Further foundation encoders, frozen Eleven further EEG foundation models — sixteen encoder checkpoints — were swapped into the published LaBraM and CBraMod frozen recipe, with the same windows, people, folds, ridge head and scoring; five of these checkpoints have a random-initialization control (above), and the others have none. EEGPT sits in the sleep and BETA tables too, from its own release: it ran in the same runs as the published LaBraM and CBraMod rows with only the encoder swapped, and its random-initialization control is in the sections above. A row is above or below another only where their 95% intervals do not overlap; with 126 frozen cells from the further encoders and EEGPT’s eight more, an occasional non-overlap is expected by chance. ### Sleep staging: four new encoders above every published row On the core matrix’s sleep protocol (EESM19 scalp subset, 20 people, six channels, chance 20.0%), REVE Large, ST-EEGFormer Base and ST-EEGFormer Large lie entirely above every published row — spectral ridge, the published non-foundation row with the highest point estimate (72.2%, 70.7%–73.8%), the frozen LaBraM and CBraMod rows, and EEGNet. So does EEGPT, by a narrow margin: the lower end of its interval, 73.9%, clears the upper end of spectral ridge’s by 0.10 pp. Sleep is the only one of the seven scored protocols where a new row clears the published non-foundation row with the highest point estimate. The other new rows overlap the published ones or lie below them. **Published rows on sleep staging (core matrix):** Spectral ridge 72.2% (70.7%–73.8%) · LaBraM 71.1% (69.0%–73.2%) · CBraMod 71.3% (68.7%–73.7%) · EEGNet 56.5% (54.9%–58.3%) Sleep staging · frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval. | Encoder | Balanced accuracy | Against LaBraM, CBraMod and spectral ridge | | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | 74.6% (72.9%–76.4%) | Overlaps all three | | [REVE Large](https://bci.report/methods/reve/) | 76.6% (75.2%–78.0%) | Above all three | | [LUNA Base](https://bci.report/methods/luna/) | 67.5% (65.3%–69.5%) | Overlaps LaBraM and CBraMod; below spectral ridge | | [LUNA Large](https://bci.report/methods/luna/) | 72.4% (70.4%–74.0%) | Overlaps all three | | [BrainOmni Base](https://bci.report/methods/brainomni/) | 67.7% (65.8%–69.5%) | Overlaps LaBraM and CBraMod; below spectral ridge | | [CodeBrain](https://bci.report/methods/codebrain/) | 70.7% (68.8%–72.7%) | Overlaps all three | | [EEGMamba](https://bci.report/methods/eegmamba/) | 64.1% (61.6%–66.4%) | Below all three | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | 77.3% (75.6%–78.8%) | Above all three | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | 79.6% (78.2%–81.0%) | Above all three | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | 67.9% (65.4%–70.4%) | Overlaps LaBraM and CBraMod; below spectral ridge | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | 71.0% (69.5%–72.6%) | Overlaps all three | | [SingLEM](https://bci.report/methods/singlem/) | 35.3% (33.7%–37.0%) | Below all three | | [ZUNA 1.1](https://bci.report/methods/zuna/) — Model card: research use only, not for diagnosis or clinical use. | 60.1% (57.8%–62.3%) Exposure unknown | Below all three | | [EEGPT](https://bci.report/methods/eegpt/) — Same runs as the published rows | 75.5% (73.9%–77.1%) | Above LaBraM (intervals 0.67 pp apart), CBraMod (intervals 0.17 pp apart) and spectral ridge (intervals 0.10 pp apart) | Against the published rows: LaBraM, CBraMod and spectral ridge, the non-foundation row of experiments.json with the highest point estimate, by whether the 95% intervals overlap. Unless a row is marked, its dataset is not in the model authors’ published pretraining list (checked 2026-10-04); each model’s sources are on its method page. EEGPT’s row gives the gap between the two intervals beside each relation; its authors’ list was checked on 2026-10-12. ### BETA: above CBraMod, overlapping or below training-free CCA On BETA’s 40-target SSVEP with eight electrodes (70 people, chance 2.5%), REVE Large and eeg-fm-masking MAE (r=9 cm, L=2) lie above the published frozen CBraMod row (33.7%), as do the three masking-ablation siblings shown below. No new row lies above standard CCA (63.1%), which needs no training at all. ST-EEGFormer’s BETA cells are in its authors’ pretraining list, and they still lie below CBraMod. **Published rows on BETA at eight electrodes (core matrix):** Standard CCA 63.1% (57.2%–69.0%) · Spectral ridge 50.7% (45.6%–55.8%) · LaBraM 10.8% (9.5%–12.2%) · CBraMod 33.7% (29.8%–37.6%) · EEGNet 55.8% (50.1%–61.3%) SSVEP, BETA, eight electrodes · frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval. | Encoder | Balanced accuracy | Against LaBraM, CBraMod and standard CCA | | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | 37.5% (33.3%–41.9%) | Above LaBraM; overlaps CBraMod; below standard CCA | | [REVE Large](https://bci.report/methods/reve/) | 52.7% (47.7%–57.8%) | Above LaBraM and CBraMod; overlaps standard CCA | | [LUNA Base](https://bci.report/methods/luna/) | 8.6% (7.7%–9.5%) | Below all three | | [LUNA Large](https://bci.report/methods/luna/) | 10.0% (8.9%–11.1%) | Overlaps LaBraM; below CBraMod and standard CCA | | [BrainOmni Base](https://bci.report/methods/brainomni/) | 21.7% (18.6%–25.0%) | Above LaBraM; below CBraMod and standard CCA | | [CodeBrain](https://bci.report/methods/codebrain/) | 11.9% (10.5%–13.3%) | Overlaps LaBraM; below CBraMod and standard CCA | | [EEGMamba](https://bci.report/methods/eegmamba/) | 40.2% (35.7%–44.7%) | Above LaBraM; overlaps CBraMod; below standard CCA | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | 17.1% (15.5%–18.8%) in the authors' published pretraining list | Above LaBraM; below CBraMod and standard CCA | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | 22.6% (20.1%–25.3%) in the authors' published pretraining list | Above LaBraM; below CBraMod and standard CCA | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | 54.8% (49.4%–60.0%) | Above LaBraM and CBraMod; overlaps standard CCA | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | 18.4% (16.3%–20.5%) | Above LaBraM; below CBraMod and standard CCA | | [SingLEM](https://bci.report/methods/singlem/) | 26.6% (23.7%–29.5%) | Above LaBraM; below CBraMod and standard CCA | | [ZUNA 1.1](https://bci.report/methods/zuna/) — Model card: research use only, not for diagnosis or clinical use. | 11.9% (10.8%–13.0%) Exposure unknown | Overlaps LaBraM; below CBraMod and standard CCA | | [EEGPT](https://bci.report/methods/eegpt/) — Same runs as the published rows | 20.3% (18.2%–22.6%) | Above LaBraM (intervals 5.97 pp apart); below CBraMod (intervals 7.24 pp apart) and standard CCA (intervals 34.64 pp apart) | Against the published rows: LaBraM, CBraMod and standard CCA, the non-foundation row of experiments.json with the highest point estimate (its interval overlaps EEGNet’s), by whether the 95% intervals overlap. Unless a row is marked, its dataset is not in the model authors’ published pretraining list (checked 2026-10-04); each model’s sources are on its method page. EEGPT’s row gives the gap between the two intervals beside each relation; its authors’ list was checked on 2026-10-12. **EEGPT on BETA:** PO5 and PO6 use channel embeddings that EEGPT’s pretraining never trained: they are two of the four names in its 62-name channel dictionary (AF7, AF8, PO5, PO6) that its 58-electrode pretraining montage leaves out. The other six channels here, and all four of the 4-channel subset, use trained embeddings. For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated. ### EEGMAT: frozen readouts and LoRA disagree Nine of the new encoders were also adapted on EEGMAT, mental arithmetic against rest (36 people, 5 participant-disjoint folds, three seeds, chance 50.0%), with the recipe of the published LaBraM adaptation: a head trained on the frozen encoder against rank-4 LoRA with the same head, five epochs, no tuning. LoRA’s budget differs by architecture, from 32,768 to 428,032 added trainable parameters. The ridge readout of the frozen-probe table, the trained head and LoRA do not put the encoders in the same order. LoRA’s change excludes zero for REVE Base, REVE Large, LUNA Base, SingLEM and ZUNA 1.1; it includes zero for BrainOmni Base, CodeBrain, EEGMamba and ST-EEGFormer Base. The clearest case is SingLEM: both frozen readouts reach chance (51.5% with the ridge head, 50.9% with the trained head), while LoRA reached 60.6%. EEGMAT · balanced accuracy with its descriptive 95% interval · the paired change over the same people and folds, with how many people moved which way · trainable parameters of the head alone and of the LoRA arm. | Encoder | Frozen, ridge head (frozen-probe table) | Frozen, trained head | LoRA rank 4 + head | LoRA minus trained head | Trainable parameters · head · LoRA arm | | --- | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) — LaBraM, already published | 64.6% (60.5%–68.6%) | 56.6% (54.2%–59.0%) | 64.1% (60.2%–67.8%) | +7.5 pp (+4.8 to +10.1 pp) — 30 improved · 6 worse · 0 unchanged The interval excludes zero. | 402 · 38,802 | | [REVE Base](https://bci.report/methods/reve/) | 68.8% (64.1%–73.2%) | 64.0% (60.4%–67.7%) | 66.3% (61.9%–70.7%) | +2.3 pp (+0.1 to +4.6 pp) — 21 improved · 13 worse · 2 unchanged The interval excludes zero. | 1,026 · 181,250 | | [REVE Large](https://bci.report/methods/reve/) | 69.0% (63.3%–74.7%) | 65.6% (61.1%–70.1%) | 69.2% (63.4%–74.9%) | +3.6 pp (+0.5 to +6.8 pp) — 21 improved · 13 worse · 2 unchanged The interval excludes zero. | 2,434 · 430,466 | | [LUNA Base](https://bci.report/methods/luna/) | 60.7% (56.9%–64.4%) | 58.2% (54.5%–61.7%) | 60.0% (55.9%–63.9%) | +1.8 pp (+0.1 to +3.3 pp) — 23 improved · 11 worse · 2 unchanged The interval excludes zero. | 514 · 37,378 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | 61.6% (58.4%–64.9%) | 64.5% (59.6%–69.2%) | 64.8% (59.6%–69.7%) | +0.3 pp (−1.4 to +1.8 pp) — 18 improved · 14 worse · 4 unchanged The interval includes zero: no change is established. | 16,386 · 106,498 | | [CodeBrain](https://bci.report/methods/codebrain/) | 63.0% (59.5%–66.5%) | 66.5% (62.3%–70.6%) | 67.1% (62.6%–71.6%) | +0.6 pp (−0.7 to +1.9 pp) — 18 improved · 14 worse · 4 unchanged The interval includes zero: no change is established. | 7,602 · 58,802 | | [EEGMamba](https://bci.report/methods/eegmamba/) | 60.3% (57.0%–63.4%) | 60.4% (55.7%–65.2%) | 60.8% (56.3%–65.3%) | +0.4 pp (−0.7 to +1.5 pp) — 18 improved · 14 worse · 4 unchanged The interval includes zero: no change is established. | 7,602 · 90,930 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | 64.5% (59.0%–70.0%) | 63.2% (59.7%–66.7%) | 65.1% (61.0%–69.4%) | +1.9 pp (−0.2 to +4.1 pp) — 20 improved · 16 worse · 0 unchanged The interval includes zero: no change is established. | 1,538 · 148,994 | | [SingLEM](https://bci.report/methods/singlem/) | 51.5% (49.2%–53.8%) Interval reaches chance level | 50.9% (49.5%–52.3%) Interval reaches chance level | 60.6% (56.7%–64.4%) | +9.7 pp (+6.1 to +13.3 pp) — 30 improved · 6 worse · 0 unchanged The interval excludes zero. | 1,218 · 33,986 | | [ZUNA 1.1](https://bci.report/methods/zuna/) Exposure unknown — Model card: research use only, not for diagnosis or clinical use. | 53.3% (49.5%–57.1%) Interval reaches chance level | 51.2% (49.8%–52.7%) Interval reaches chance level | 56.9% (53.7%–60.1%) | +5.7 pp (+2.5 to +8.7 pp) — 28 improved · 7 worse · 1 unchanged The interval excludes zero. | 1,218 · 296,130 | **The others were not adapted:** LUNA Large (not run by design: LUNA Large is frozen probes only in this evaluation’s specification); ST-EEGFormer Large (not run by design: ST-EEGFormer Large is frozen probes only); eeg-fm-masking MAE (r=9 cm, L=2), eeg-fm-masking JEPA (r=9 cm, L=2), eeg-fm-masking MAE (r=one channel, L=1) and eeg-fm-masking JEPA (r=one channel, L=1) (not run by design: eeg-fm-masking is a frozen-probe masking ablation); ERP-FM Base (not run by design: ERP-FM is frozen probes only and EEGMAT arithmetic is outside the ERP design). ### REVE Base and Large: a size comparison REVE Large’s point estimate is at or above REVE Base’s on all seven scored protocols; their intervals do not overlap on BETA at eight electrodes, BETA at four electrodes, the P300 target ERP and the semantic target ERP. Both checkpoints were scored on the same people, but each interval is marginal and no paired test was run, so a gap is a description, not an established effect of size. REVE Base and Large on every scored protocol: the table Frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval · Large minus Base as the export states it, and whether the two intervals overlap. | Protocol | REVE Base | REVE Large | Large minus Base | Intervals | | --- | --- | --- | --- | --- | | [Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) — 10 people | 66.2% (61.8%–70.6%) | 70.2% (64.0%–76.1%) | +4.0 pp | Overlap | | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) — 70 people | 37.5% (33.3%–41.9%) | 52.7% (47.7%–57.8%) | +15.2 pp | **Do not overlap** | | [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) — 70 people | 25.5% (22.4%–28.8%) | 46.4% (41.7%–51.2%) | +20.8 pp | **Do not overlap** | | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) — 36 people | 68.8% (64.1%–73.2%) | 69.0% (63.3%–74.7%) | +0.2 pp | Overlap | | [P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) — 21 people | 50.5% (48.5%–52.3%) Interval reaches chance level | 54.6% (52.7%–56.5%) | +4.1 pp | **Do not overlap** | | [Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) — 30 people | 53.7% (51.8%–55.4%) | 59.7% (57.5%–61.8%) | +6.0 pp | **Do not overlap** | | [Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) — 20 people | 74.6% (72.9%–76.4%) | 76.6% (75.2%–78.0%) | +2.0 pp | Overlap | The other two size pairs do not show this pattern: LUNA’s Base and Large intervals separate only on sleep staging, ST-EEGFormer’s only on BETA at eight electrodes and BETA at four electrodes, where its cells are in the authors’ pretraining list. ### Masking ablation: not separable on six of seven protocols eeg-fm-masking released four checkpoints of one study: masking framework (MAE or JEPA) crossed with mask geometry (the paper-recommended r = 9 cm with L = 2, or a single channel with L = 1), with the same architecture, corpus and recipe. The paper-recommended MAE checkpoint is a matrix row in the tables above; the other three are its ablation. On six of the seven scored protocols every pair of the four overlaps. On sleep staging one pair separates — JEPA at r = 9 cm against MAE on one channel — and it changes the framework and the geometry at once, so it isolates neither. The four eeg-fm-masking checkpoints on every scored protocol: the table Frozen encoder and ridge head · balanced accuracy with its descriptive 95% interval. | Protocol | MAE · r = 9 cm, L = 2 (matrix row) | JEPA · r = 9 cm, L = 2 | MAE · one channel, L = 1 | JEPA · one channel, L = 1 | Pairs whose intervals separate | | --- | --- | --- | --- | --- | --- | | [Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) | 68.1% (64.3%–72.3%) | 63.2% (58.2%–67.9%) | 64.9% (60.3%–69.4%) | 63.3% (59.0%–67.4%) | None: every pair overlaps | | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) | 54.8% (49.4%–60.0%) | 49.6% (44.5%–54.8%) | 53.4% (48.1%–58.6%) | 51.0% (45.8%–56.1%) | None: every pair overlaps | | [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) | 47.1% (42.1%–52.1%) | 43.4% (38.7%–48.3%) | 46.7% (41.9%–51.5%) | 45.4% (40.6%–50.2%) | None: every pair overlaps | | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) | 63.6% (59.1%–68.1%) | 61.6% (56.4%–66.5%) | 64.3% (60.2%–68.2%) | 62.5% (58.5%–66.2%) | None: every pair overlaps | | [P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) | 49.5% (48.1%–51.0%) At or below chance level | 51.6% (50.4%–52.7%) | 51.1% (49.2%–53.0%) Interval reaches chance level | 50.0% (48.7%–51.5%) Interval reaches chance level | None: every pair overlaps | | [Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) | 54.8% (52.7%–56.9%) | 55.8% (53.9%–57.6%) | 54.9% (53.4%–56.4%) | 55.0% (53.4%–56.6%) | None: every pair overlaps | | [Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) | 67.9% (65.4%–70.4%) | 70.3% (68.0%–72.7%) | 65.2% (63.0%–67.3%) | 68.3% (65.7%–70.8%) | **JEPA at r = 9 cm vs. MAE on one channel** | ### Limits of the further encoders ### One fixed recipe, not a ceiling The frozen probe is the published linear head with only the encoder swapped; the adaptation is one untuned five-epoch recipe on one task. ### Small cohorts, descriptive intervals Motor imagery has 10 people, P300 21, sleep staging 20. The intervals are participant bootstraps that ignore cross-validation dependence and carry no multiplicity correction. ### Family order Rows keep the export’s family order; the same people under two encoders are not tested as a pair. ### Method comparisons Balanced or class-matched designs (P300, semantic target, motor imagery against rest, sampled sleep epochs) compare methods; they are not detection, false-alarm or latency estimates for real use. ### New people, same task and setup The folds hold out people. Nothing here is cross-day, cross-device or cross-dataset evidence, and nothing is a clinical claim. ### Short windows Every encoder sees the published 1–2 s windows (sleep as 15 two-second segments), usually shorter than its pretraining context; each model’s input handling is in its row footnote on the protocol pages. ### Pretraining exposure: the authors’ lists BETA is in ST-EEGFormer’s published pretraining list and TMNRED in SingLEM’s; ZUNA 1.1 publishes none, so every ZUNA 1.1 cell is unknown; EEGMamba’s list is read from its official code, with medium confidence. Every other cell, LaBraM’s and CBraMod’s included, is not in the authors’ published pretraining list (checked 2026-10-04), and neither are EEGPT’s (checked 2026-10-12) — not proof that the recordings were never seen. ### No endorsement No model’s authors endorse these results. No adapted weights or LoRA deltas were kept or shared, and each model’s weights licence travels with its rows. Timings were measured on a shared GPU and are not published. ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. Every row, with its footnote and exposure source, is on its protocol page ([Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9), [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9), [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9), [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9), [P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9), [Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9), [Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9)); the adaptation arms and the comparisons above are in [foundation-models-update.json](https://bci.report/data/foundation-models-update.json). Release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004). EEGPT’s rows are in their own section of the same protocol pages, from [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json), release [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013). Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder were run in the same frozen recipe as the encoders above: the same windows, people, folds, ridge head and scoring, with no random-initialization control. ### The Quark ladder: B, L, H and G Read by the same frozen linear probe on six datasets, Quark-L minus Quark-B averaged −0.93 pp (95% interval −1.75 to −0.09), and Quark-H minus Quark-L averaged +0.03 pp (−0.90 to +0.92). Under this probe, among Quark-B, Quark-L and Quark-H, the larger released encoders did not reach a higher average balanced accuracy on these protocols. Quark-G minus Quark-H, with both encoders run on the same GPU, averaged −2.33 pp over the six datasets, dataset-weighted (95% interval −3.30 to −1.32); the probe's input width doubles from 800 to 1600 under one fixed penalty, so this step includes that change. One fixed ridge probe (α = 100 on standardised training-fold features) reads every encoder; the feature width grows from 200 to 400 to 800 to 1600, so each step, the Quark-G step included, includes the probe's width under one penalty. Each Quark size, Quark-G included, comes with its own pretraining recipe; the ladder compares the encoders as released. Each of the six datasets weighs one sixth; the two BETA montages, the same 70 people, one twelfth each. The values are balanced-accuracy points across 2-, 5- and 40-class protocols, not normalised effects. The intervals resample people within these datasets, not the choice of datasets, and the three step summaries, prespecified in two protocols, carry no multiplicity correction. Balanced accuracy of each Quark encoder, frozen, by protocol, with its 95% interval. | Encoder | Motor imagery & rest | SSVEP · 8 channels | SSVEP · 4 channels | Arithmetic & rest | P300 target ERP | Semantic target ERP | Sleep staging | | --- | --- | --- | --- | --- | --- | --- | --- | | Quark-B | 61.9% — 59.1%–64.4% | 27.4% — 23.6%–31.5% | 27.8% — 24.0%–31.9% | 66.0% — 62.0%–70.0% | 49.6% — 48.5%–50.6% · interval includes chance | 55.4% — 53.1%–57.8% | 72.1% — 70.3%–74.0% | | Quark-L | 59.6% — 56.3%–62.9% | 27.1% — 23.3%–31.0% | 28.0% — 24.0%–32.2% | 63.8% — 59.8%–67.8% | 51.0% — 49.4%–52.6% · interval includes chance | 54.8% — 53.0%–56.6% | 70.3% — 68.0%–72.5% | | Quark-H | 60.2% — 57.4%–62.9% | 23.7% — 20.2%–27.3% | 26.2% — 22.3%–30.3% | 64.5% — 60.6%–68.3% | 52.6% — 51.1%–54.1% | 52.2% — 50.3%–53.9% | 72.9% — 71.5%–74.2% | | Quark-G † | 58.7% — 57.3%–60.1% | 25.0% — 21.3%–28.8% | 26.8% — 22.8%–31.1% | 59.4% — 56.9%–61.7% | 50.3% — 48.5%–52.1% · interval includes chance | 51.8% — 50.1%–53.4% | 67.3% — 65.6%–68.9% | Per protocol: the paired differences Paired participant bootstrap, 10,000 draws; difference in balanced-accuracy points. | Protocol | People | Difference | 95% interval | | --- | --- | --- | --- | | Quark-L minus Quark-B | | | | | Motor imagery & rest | 10 | −2.33 pp | −4.83 to −0.25 pp | | SSVEP · 8 channels | 70 | −0.39 pp | −1.35 to +0.58 pp | | SSVEP · 4 channels | 70 | +0.14 pp | −0.87 to +1.17 pp | | Arithmetic & rest | 36 | −2.13 pp | −5.05 to +0.79 pp | | P300 target ERP | 21 | +1.43 pp | −0.32 to +3.29 pp | | Semantic target ERP | 30 | −0.64 pp | −2.83 to +1.67 pp | | Sleep staging | 20 | −1.80 pp | −3.30 to −0.27 pp | | Quark-H minus Quark-L | | | | | Motor imagery & rest | 10 | +0.58 pp | −3.08 to +3.50 pp | | SSVEP · 8 channels | 70 | −3.38 pp | −4.37 to −2.45 pp | | SSVEP · 4 channels | 70 | −1.74 pp | −2.74 to −0.76 pp | | Arithmetic & rest | 36 | +0.65 pp | −1.67 to +2.83 pp | | P300 target ERP | 21 | +1.59 pp | −0.40 to +3.49 pp | | Semantic target ERP | 30 | −2.61 pp | −5.11 to −0.11 pp | | Sleep staging | 20 | +2.53 pp | +1.00 to +4.20 pp | | Quark-H minus Quark-B | | | | | Motor imagery & rest | 10 | −1.75 pp | −5.17 to +1.00 pp | | SSVEP · 8 channels | 70 | −3.78 pp | −4.78 to −2.79 pp | | SSVEP · 4 channels | 70 | −1.60 pp | −2.50 to −0.67 pp | | Arithmetic & rest | 36 | −1.48 pp | −4.35 to +1.25 pp | | P300 target ERP | 21 | +3.02 pp | +1.31 to +4.84 pp | | Semantic target ERP | 30 | −3.25 pp | −5.53 to −0.94 pp | | Sleep staging | 20 | +0.73 pp | −0.93 to +2.50 pp | | Quark-G minus Quark-H, both on the same GPU: paired participant bootstrap, 10,000 draws; the probe's input width doubles from 800 to 1600 under one fixed penalty.† | | | | | Motor imagery & rest | 10 | −1.50 pp | −3.92 to +1.58 pp | | SSVEP · 8 channels | 70 | +1.31 pp | +0.34 to +2.35 pp | | SSVEP · 4 channels | 70 | +0.61 pp | −0.38 to +1.57 pp | | Arithmetic & rest | 36 | −5.14 pp | −8.84 to −1.44 pp | | P300 target ERP | 21 | −2.26 pp | −4.68 to +0.08 pp | | Semantic target ERP | 30 | −0.44 pp | −2.86 to +1.94 pp | | Sleep staging | 20 | −5.60 pp | −6.87 to −4.20 pp | Per protocol, Quark-L was lower than Quark-B on motor imagery and rest and on sleep staging; Quark-H was lower than Quark-L on 8-channel BETA, on 4-channel BETA and on the semantic target, and higher on sleep staging. The other intervals of these two steps include zero. If no true difference existed, about one in twenty of the per-protocol intervals would exclude zero by chance; the near-zero average of Quark-H minus Quark-L is made of differences in both directions, and the two BETA montages are the same people. Per protocol, Quark-G was lower than Quark-H on arithmetic and rest and on sleep staging and higher on 8-channel BETA, each step including the probe's input width doubling from 800 to 1600 under one fixed penalty; the other 4 intervals include zero. If no true difference existed, about 0.35 of the 7 intervals would exclude zero by chance; the rows for sleep staging and for arithmetic and rest carry most of the average, and the two BETA montages are the same people. † Quark-G and the Quark-H baseline of its step ran on Local Ubuntu / CUDA; the Quark-H cells shown ran on Local Ubuntu / CPU. The GPU re-run reproduces them except two window predictions on beta-4ch, where the step reads +0.61 pp against the GPU baseline (+0.60 pp against the cell shown) and, like every Quark-G step, includes the probe's input width doubling from 800 to 1600 under one fixed penalty. ### Atom-L and Molecule-L: two other collections The Large encoders of LaBraM at Scale's Atom and Molecule collections are read the same way. Their pretraining lists include five of the core datasets, so their cells on motor imagery and rest, idle, arithmetic and rest, P300 target and semantic target carry the flag "its pretraining corpus includes this protocol's dataset". Atom and Molecule differ from Quark in corpus size, composition and input normalisation, so these rows stand on their own. Balanced accuracy by protocol, with its 95% interval; a flagged cell’s dataset is in that collection’s pretraining list. | Encoder | Motor imagery & rest | SSVEP · 8 channels | SSVEP · 4 channels | Arithmetic & rest | P300 target ERP | Semantic target ERP | Sleep staging | | --- | --- | --- | --- | --- | --- | --- | --- | | Atom-L | 56.9% — 53.3%–61.0% — its pretraining corpus includes this protocol's dataset | 25.8% — 22.4%–29.4% | 28.0% — 24.3%–31.8% | 67.5% — 63.7%–71.3% — its pretraining corpus includes this protocol's dataset | 53.2% — 51.5%–54.9% — its pretraining corpus includes this protocol's dataset | 52.8% — 51.3%–54.3% — its pretraining corpus includes this protocol's dataset | 72.7% — 70.6%–74.4% | | Molecule-L | 59.6% — 57.2%–61.4% — its pretraining corpus includes this protocol's dataset | 26.6% — 22.9%–30.5% | 28.9% — 25.1%–33.0% | 63.9% — 59.9%–68.0% — its pretraining corpus includes this protocol's dataset | 51.4% — 49.2%–53.5% · interval includes chance — its pretraining corpus includes this protocol's dataset | 54.7% — 52.9%–56.5% — its pretraining corpus includes this protocol's dataset | 71.9% — 70.2%–73.4% | ### REVE Small, an early checkpoint (epoch 15 of 200) REVE Small's encoder was released at epoch 15 of a 200-epoch schedule; every REVE Small value carries the label "early checkpoint (epoch 15 of 200)". REVE Small reads its input as the REVE Base and Large rows do: each segment's per-channel mean is removed before dividing by 100; no sensitivity run was made for Small. REVE Small's card does not restate its corpus; its exposure is read from the REVE corpus list (checked 2026-10-04), which includes none of the core datasets. Balanced accuracy by protocol, with its 95% interval. | Encoder | Motor imagery & rest | SSVEP · 8 channels | SSVEP · 4 channels | Arithmetic & rest | P300 target ERP | Semantic target ERP | Sleep staging | | --- | --- | --- | --- | --- | --- | --- | --- | | REVE Small — early checkpoint (epoch 15 of 200) | 61.5% — 58.3%–64.8% | 39.3% — 35.2%–43.6% | 35.9% — 32.0%–40.0% | 65.2% — 60.5%–70.0% | 54.3% — 52.9%–55.9% | 56.7% — 54.5%–59.0% | 64.5% — 62.5%–66.6% | ### Idle & command: trial counts Cohort size n = 4 (a protocol fact): idle trials with a false trigger, command trials detected within 3 s, and the window metrics; no contrast. | Encoder | Idle false triggers | Commands detected ≤3 s | Mean window balanced accuracy | Mean window AUROC | | --- | --- | --- | --- | --- | | Quark-B | 4 / 60 | 15 / 60 | 60.8% | 0.724 | | Quark-L | 2 / 60 | 13 / 60 | 61.1% | 0.655 | | Quark-H | 2 / 60 | 15 / 60 | 62.8% | 0.746 | | Quark-G | 2 / 60 | 10 / 60 | 57.2% | 0.722 | | Atom-L — its pretraining corpus includes this protocol's dataset | 2 / 60 | 20 / 60 | 66.1% | 0.721 | | Molecule-L — its pretraining corpus includes this protocol's dataset | 1 / 60 | 10 / 60 | 60.6% | 0.723 | | REVE Small — early checkpoint (epoch 15 of 200) | 1 / 60 | 8 / 60 | 54.4% | 0.680 | Idle is a within-person protocol: each row gives the cohort size, trial counts and window metrics, with no contrast. LaBraM at Scale's pooled features are global on idle; the published LaBraM idle row kept the channel axis. ### How these rows were read - The published 250 Hz microvolt windows, resampled from 250 to 200 Hz (resample_poly 4, 5); 1 s and 2 s windows are one or two patches; sleep is the mean of the pooled features of 15 two-second segments; recorded channels only, with no padding or interpolation. - Ridge, α = 100 (Cholesky), on train-fold-standardised features, class-balanced except on BETA; idle: a per-person logistic head with a calibration-only threshold and a trigger after two consecutive decisions. - Participant-held-out: 5 folds, 7 on BETA. - Cells: a descriptive participant bootstrap of the equal-person mean, 10,000 draws. Paired differences: a paired participant bootstrap, 10,000 draws, one draw matrix per dataset, with BETA's two montages resampled jointly. - Extraction and heads are deterministic: each cell has one value. - Published cells re-extracted in this recipe, each matching its published value: REVE Base, arithmetic and rest; LaBraM, arithmetic and rest. - The audit re-extracted sample windows of two LaBraM at Scale encoders, Quark-B and Molecule-L, with its own code and matched them. - An export-stage check re-extracted sample windows of REVE Small on CPU and matched them. - On the GPU path, one published cell (LaBraM, arithmetic and rest) was re-extracted and matched its published value. ### Weights terms LaBraM at Scale: code and weights under MIT, ungated (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27, Quark-G included; code AlbertMentat/LaBraM-at-Scale @ faa387d9). Cite [arXiv 2611.03165](https://arxiv.org/abs/2611.03165) and, for the architecture, LaBraM ([arXiv 2405.18765](https://arxiv.org/abs/2405.18765)). REVE Small: the REVE Responsible Use License v1.0 (model version reve-small @ 3565bdcb), aggregate scientific results only; cite [arXiv 2510.21585](https://arxiv.org/abs/2510.21585). No weights are distributed. Every cell is also on its protocol page: [Motor imagery & rest](https://bci.report/protocols/mi-rest/#more-checkpoints); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#more-checkpoints); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#more-checkpoints); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#more-checkpoints); [P300 target ERP](https://bci.report/protocols/p300-target/#more-checkpoints); [Semantic target ERP](https://bci.report/protocols/semantic-target/#more-checkpoints); [Sleep staging](https://bci.report/protocols/sleep-scalp/#more-checkpoints); [Idle & command](https://bci.report/protocols/idle/#more-checkpoints). The cells and the ladder’s differences are in [foundation-checkpoints-update.json](https://bci.report/data/foundation-checkpoints-update.json). Release [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106). Methods & limits ## Limits of the pretraining contrasts ### Fixed versus selected heads The primary head fixes ridge alpha at 100. The sensitivity setting selects among 100, 10,000 and 1 using inner training participants only. ### Three seeds, one cohort The random baseline averages three encoder initializations. The participant bootstrap does not include unrestricted retraining uncertainty. ### Task-specific evidence MI/rest contains 10 participants and 1,200 trials; mental workload contains 36 participants and 2,160 trials. Sleep staging, in the further encoders’ controls, contains 20 participants and 3,000 epochs. ### Exposure of the two benchmarks Neither ds003810 nor EEGMAT is in LaBraM’s or CBraMod’s published pretraining list, checked on 4 October 2026, or in EEGPT’s, checked on 12 October 2026 (sources below). The lists of REVE, ST-EEGFormer, SingLEM and EEGMamba, checked for the three datasets of their random-initialization controls, are stated with those controls. ### Mental workload’s recording order EEGMAT records rest before the arithmetic task for every person, so the two classes also differ in time on task. Pretraining lists, checked 2026-10-04: [LaBraM ↗](https://arxiv.org/abs/2405.18765) · [CBraMod ↗](https://arxiv.org/abs/2412.07236) · EEGPT’s, checked 2026-10-12: [EEGPT ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [EEGPT repository ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) Frozen CBraMod with a ridge readout is also read on a later session of the same person on WBCIC-SHU, without a random-initialization control: [Does a decoder trained on an earlier session still work later?](https://bci.report/topics/later-sessions/#wbcic-shu) Frozen EEGPT with a supervised probe is read against matched classical features in one small pilot, on recorded orthosis device errors, without a random-initialization control: [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/#orthosis) A log band-power ridge (5 bands, visible channels) sits beside the Spectral ridge row on seven core protocols, with its paired difference from one frozen-encoder row on each: [how it is built and where it appears →](https://bci.report/methods/#log-band-power-ridge) A small model trained on EEGMAT alone is also set against the same model never trained, with heads fitted on part or all of the training labels: [The trained model against the same model never trained →](https://bci.report/topics/pilot-studies/#nsp-components-supplement) ### MI / rest · OpenNeuro ds003810 v2.0.2 Peterson et al. · CC0-1.0. [Dataset record ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) ### Mental workload · EEGMAT 1.0.0 Zyma et al. · Open Data Commons Attribution License 1.0. [PhysioNet record ↗](https://physionet.org/content/eegmat/1.0.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/deployment-topics.json) · schema bci-report-public-deployment-topics-v1. EEGPT’s rows: [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) · schema bci-report-eegpt-core-update-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Does pretraining help EEG foundation models like LaBraM and CBraMod?* https://bci.report/topics/does-pretraining-help/ Figures from releases [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`random-init-controls-update-20261106`](https://bci.report/releases/#random-init-controls-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/does-pretraining-help/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Can one EEG model answer several questions about the same data as well as separate models, and at what cost? Shared encoders · route 2 of the decision-research plan One shared encoder for several questions, against a separate model per question and against a head told which question it is answering, at matched data and compute: motor imagery on OpenBMI, sleep on BOAS, EESM19 as a crude replication. - [Reviewed aggregate JSON ↓](https://bci.report/data/shared-representation-update.json) - Measured on: [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/); [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) · Methods: [EEGNet](https://bci.report/methods/eegnet/); [CBraMod](https://bci.report/methods/cbramod/) ## Short answer The cheaper set-up was not shown to do as well. A shared EEGNet with a fixed linear output per question cost less than a head told which question it was answering (**3,060** against **20,116** trainable parameters on motor imagery) but was not shown to come within 2 pp of it, the margin fixed before the run: on imagery against rest, in **51** people (OpenBMI), that head scored **+1.48 pp** higher (**+0.09** to **+2.89 pp**); on five-stage sleep scoring (BOAS) its whole interval lay above 2 pp. Naming the question added nothing measurable: the same hidden layer, not told the question, was equivalent to that head within ±2 pp on both motor-imagery questions and non-inferior at 2 pp on sleep. Against a separate EEGNet per question, the small shared EEGNet trunk scored **−1.43 pp** on which hand (**−2.57** to **−0.39 pp**): a difference, and a loss of 2 pp or more cannot be ruled out; on imagery or rest and on the sleep stage the comparison was inconclusive. How the comparison works ## Four set-ups, the same questions Every question is a label the dataset already has, and the model is told by its identifier which one it is answering. The four set-ups answer the same questions from the same windows. - **A · Separate models**: one model per question. - **B-lin · Fixed heads**: one shared encoder and one linear output per question; the fewest parameters. - **B-sh · Shared hidden layer**: the shared encoder, then one hidden layer used by every question, then one output per question. - **C1 · Question-conditioned head**: B-sh, with the hidden layer adjusted for each question by a learned code (FiLM). B-lin, B-sh and C1 start from the same weights, see the same batches in the same order and train for the same number of steps. Two primary levels, three seeds each: EEGNet trained from scratch (E1), and heads trained on frozen CBraMod features (L1; the same set-ups, marked -fz). CBraMod adapted by LoRA (E2) ran with one seed, as a secondary level. P1 (C1 − B-lin) asks whether fixed heads do as well; P5 (C1 − B-sh), whether telling the head the question adds anything; P2 (B-lin − A), whether sharing one encoder costs accuracy. P4 is P1 on frozen CBraMod features. Each comparison is a paired difference in balanced accuracy, in percentage points (pp), with a 95% interval from resampling people (and, with three seeds, training runs). A difference is shown when the interval excludes zero. Against a margin of ±2 pp, fixed by the owner before any result, a comparison is equivalent when the whole interval lies inside the margin, non-inferior when the bound that matters does, and otherwise the margin is not met. “Fixed heads do as well for less” is said of a domain only if every counted question meets the margin on P1 and the fixed heads cost less. A question counts only where the fixed heads’ 95% interval lies wholly above chance + 5 pp; otherwise it is at floor and only reported. There are 21 primary comparisons and no multiplicity correction, so about one in 20 comparisons with no true difference may show one by chance. Primary · OpenBMI · EEGNet and CBraMod ## Motor imagery: imagery or rest, and which hand Two questions about each 2-second window of OpenBMI’s motor-imagery runs and rest recordings: is the person imagining a hand movement or resting (MI-A), and which hand (MI-B, offline trials only)? 31,653 windows from 51 people; each person’s two sessions sit in one fold, so every test is on new people. Both questions count. **+1.48 pp** C1 − B-lin on MI-A, EEGNet from scratch, three seeds ### Fixed heads did as well for less: not supported On imagery or rest the conditioned head scored +1.48 pp above the fixed heads (+0.09 to +2.89 pp): a gap of 2 pp or more cannot be ruled out. On which hand the two were equivalent within ±2 pp (−0.09 pp, −1.24 to +1.04 pp). The fixed heads cost less: 3,060 against 20,116 trainable parameters, and 960 against 15,744 head multiply-accumulates per window. Paired differences in balanced accuracy, in pp, with 95% intervals; 51 people, three seeds. | Comparison | Question | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | | EEGNet trained from scratch (E1) | | | | | | | P1 · Conditioned head − fixed heads — C1 − B-lin | MI-A — imagery or rest | +1.48 pp (95% interval +0.09 to +2.89 pp) | C1 higher | 2 pp margin not met | Gate passed | | P1 · Conditioned head − fixed heads — C1 − B-lin | MI-B — which hand | −0.09 pp (95% interval −1.24 to +1.04 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | MI-A — imagery or rest | +0.31 pp (95% interval −0.91 to +1.54 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | MI-B — which hand | +0.45 pp (95% interval −0.60 to +1.34 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(MI-A) | MI-A — imagery or rest | −1.00 pp (95% interval −2.21 to +0.19 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed | | P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(MI-B) | MI-B — which hand | −1.43 pp (95% interval −2.57 to −0.39 pp) | A(MI-B) higher | 2 pp margin not met | Gate passed | | Heads on frozen CBraMod features (L1) | | | | | | | P4 · Conditioned head − fixed heads — C1-fz − B-lin-fz-sgd | MI-A — imagery or rest | +1.54 pp (95% interval +0.86 to +2.30 pp) | C1-fz higher | 2 pp margin not met | Gate passed | | P4 · Conditioned head − fixed heads — C1-fz − B-lin-fz-sgd | MI-B — which hand | +1.09 pp (95% interval +0.08 to +2.12 pp) | C1-fz higher | 2 pp margin not met — not read: at floor | At floor: reported, not counted | | S1 · Conditioned head − the same layer, not told the question · secondary — C1-fz − B-sh-fz | MI-A — imagery or rest | −0.12 pp (95% interval −0.64 to +0.33 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | S1 · Conditioned head − the same layer, not told the question · secondary — C1-fz − B-sh-fz | MI-B — which hand | +0.43 pp (95% interval −0.39 to +1.21 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted | Telling the head the question added nothing measurable: C1 − B-sh was equivalent within ±2 pp on both questions from scratch, and on imagery or rest on frozen features. Where the conditioned head’s lift over the fixed heads comes from remains open: the same layer not told the question, against the fixed heads (B-sh − B-lin, secondary S2), was +1.17 pp on imagery or rest (−0.09 to +2.54 pp), inconclusive at this sample size. Against a separate EEGNet per question, sharing one small trunk (2,096 parameters) scored −1.43 pp on which hand (−2.57 to −0.39 pp): a difference, and a cost of 2 pp or more cannot be ruled out; on imagery or rest it was inconclusive. On frozen CBraMod features the conditioned head was again higher on imagery or rest (+1.54 pp, +0.86 to +2.30 pp); which hand is at floor there. Balanced accuracy of each set-up: the mean over 51 people with its 95% interval. Chance is 50.0%; the gate’s floor is 55.0%. The paired differences above are the comparison; one set-up is above another here only where intervals do not overlap. | Set-up | MI-A · imagery or rest | MI-B · which hand | | --- | --- | --- | | EEGNet trained from scratch (E1) | | | | Fixed heads — B-lin | 73.2% (95% interval 70.2%–76.2%) | 67.5% (95% interval 64.6%–70.5%) | | Shared hidden layer — B-sh | 74.3% (95% interval 71.3%–77.3%) | 67.0% (95% interval 64.1%–69.9%) | | Question-conditioned head — C1 | 74.6% (95% interval 71.9%–77.5%) | 67.4% (95% interval 64.4%–70.5%) | | Separate models — A | 74.2% (95% interval 71.3%–76.9%) | 69.0% (95% interval 65.9%–72.0%) | | Frozen CBraMod features (L1) | | | | Fixed heads — B-lin-fz-sgd | 69.2% (95% interval 67.0%–71.5%) | 56.1% (95% interval 54.8%–57.4%) | | Shared hidden layer — B-sh-fz | 70.9% (95% interval 68.4%–73.4%) | 56.7% (95% interval 55.4%–58.2%) | | Question-conditioned head — C1-fz | 70.8% (95% interval 68.4%–73.2%) | 57.2% (95% interval 55.8%–58.6%) | What each set-up costs on motor imagery (EEGNet, E1). Step times were measured once, interleaved across set-ups on a shared GPU, and are indicative: in the cost rule a time difference counts only beyond 1.2×; the other columns are exact. | Set-up | Trainable parameters | Encoder passes per window | Head multiply-accumulates per window | Training steps per fit | Time per step | | --- | --- | --- | --- | --- | --- | | Fixed heads — B-lin | 3,060 (encoder 2,096 · heads 964 · conditioning 0) | 1 | 960 | 7,928 | 7.71 ms | | Shared hidden layer — B-sh | 18,402 (encoder 2,096 · heads 16,306 · conditioning 0) | 1 | 16,104 | 7,928 | 7.80 ms | | Question-conditioned head — C1 | 20,116 (encoder 2,096 · heads 15,812 · conditioning 2,208) | 1 | 15,744 | 7,928 | 7.89 ms | | Separate models — A | 5,156 (one EEGNet per question) | 2 | 960 | 10,492 | 7.55 ms | ### What the canary and the gate set aside Before any fit, a ridge classifier on each window’s per-channel mean and log-variance, trained and tested within the first fold’s 40 training people, checked whether a question could be answered from those statistics alone; it demotes a question at 85.0% or more. It gave MI-A 66.1%, MI-B 55.9%, MI-C 60.1%, MI-D 47.3%, so none was demoted. The gate set one entry aside: on frozen CBraMod features the fixed heads scored 56.1% (54.8%–57.4%) on which hand; its interval does not lie wholly above the 55.0% floor, so that question is at floor there. Primary · BOAS · EEGNet and CBraMod ## Sleep: the stage, the next change and the half of the night Three questions about each 30-second epoch of the six polysomnography EEG channels, scored against the human consensus: the sleep stage (SL-A), whether the next epoch’s stage differs (SL-E), and the first or second half of the night (SL-F). 119,725 epochs from 100 people; all nights of a person sit in one fold. Stages at their natural mix: W 16.0%, N1 3.7%, N2 60.3%, N3 4.4%, REM 15.7%; the stage changes at the next epoch in 8.6% of epochs. SL-E and SL-F are largely predictable from the current stage: read off the true current stage alone, they reach 61.9% and 59.2%. **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. **+6.96 pp** C1 − B-lin on SL-A, EEGNet from scratch, three seeds ### Fixed heads did as well for less: not supported Only the sleep stage counts: on EEGNet the fixed heads are at floor on the next change (54.4%, 53.5%–55.4%) and on the half of the night (54.2%, 53.0%–55.6%). On the sleep stage the conditioned head scored +6.96 pp above the fixed heads (+4.92 to +9.09 pp): the whole interval lies above +2 pp. The fixed heads were cheaper: 34,905 against 243,817 trainable parameters, 33,696 against 240,384 head multiply-accumulates per epoch, and a shorter step. Paired differences in balanced accuracy, in pp, with 95% intervals; 100 people, three seeds. | Comparison | Question | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | | EEGNet trained from scratch (E1) | | | | | | | P1 · Conditioned head − fixed heads — C1 − B-lin | SL-A — sleep stage, five stages | +6.96 pp (95% interval +4.92 to +9.09 pp) | C1 higher | 2 pp margin not met | Gate passed | | P1 · Conditioned head − fixed heads — C1 − B-lin | SL-E — next epoch’s stage differs | +3.02 pp (95% interval +1.55 to +4.35 pp) | C1 higher | 2 pp margin not met — not read: at floor | At floor: reported, not counted | | P1 · Conditioned head − fixed heads — C1 − B-lin | SL-F — first or second half of the night | +1.42 pp (95% interval −0.25 to +2.90 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size — not read: at floor | At floor: reported, not counted | | P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | SL-A — sleep stage, five stages | −1.47 pp (95% interval −3.79 to +0.94 pp) | No difference shown | B-sh non-inferior at 2 pp | Gate passed | | P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | SL-E — next epoch’s stage differs | −1.41 pp (95% interval −2.40 to −0.50 pp) | B-sh higher | B-sh non-inferior at 2 pp — not read: at floor | At floor: reported, not counted | | P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | SL-F — first or second half of the night | −0.14 pp (95% interval −1.41 to +1.15 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted | | P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(SL-A) | SL-A — sleep stage, five stages | −0.84 pp (95% interval −2.73 to +0.91 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed | | P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(SL-E) | SL-E — next epoch’s stage differs | +1.20 pp (95% interval +0.07 to +2.44 pp) | B-lin higher | B-lin non-inferior at 2 pp — not read: at floor | At floor: reported, not counted | | P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(SL-F) | SL-F — first or second half of the night | +2.01 pp (95% interval +0.54 to +3.89 pp) | B-lin higher | B-lin non-inferior at 2 pp — not read: at floor | At floor: reported, not counted | | Heads on frozen CBraMod features (L1) | | | | | | | P4 · Conditioned head − fixed heads — C1-fz − B-lin-fz-sgd | SL-A — sleep stage, five stages | +0.19 pp (95% interval −0.74 to +1.14 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | P4 · Conditioned head − fixed heads — C1-fz − B-lin-fz-sgd | SL-E — next epoch’s stage differs | +1.75 pp (95% interval +0.92 to +2.63 pp) | C1-fz higher | 2 pp margin not met | Gate passed | | P4 · Conditioned head − fixed heads — C1-fz − B-lin-fz-sgd | SL-F — first or second half of the night | +0.77 pp (95% interval −0.14 to +1.67 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | S1 · Conditioned head − the same layer, not told the question · secondary — C1-fz − B-sh-fz | SL-A — sleep stage, five stages | −0.53 pp (95% interval −1.10 to +0.01 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | S1 · Conditioned head − the same layer, not told the question · secondary — C1-fz − B-sh-fz | SL-E — next epoch’s stage differs | −0.22 pp (95% interval −0.76 to +0.30 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | | S1 · Conditioned head − the same layer, not told the question · secondary — C1-fz − B-sh-fz | SL-F — first or second half of the night | +0.05 pp (95% interval −0.57 to +0.70 pp) | No difference shown | Equivalent within ±2 pp | Gate passed | On the sleep stage the gain is the hidden layer’s: the same layer not told the question scored +8.43 pp above the fixed heads (B-sh − B-lin, secondary S2; +5.83 to +11.12 pp), and C1 − B-sh was −1.47 pp (−3.79 to +0.94 pp), B-sh non-inferior at 2 pp; on frozen CBraMod features C1 − B-sh was equivalent within ±2 pp on all three questions. Sharing the trunk, against separate models, was inconclusive on the sleep stage. On frozen CBraMod features the fixed heads were equivalent within ±2 pp of the conditioned head on the sleep stage and the half of the night; on the next change the conditioned head was +1.75 pp higher (+0.92 to +2.63 pp). Five-stage accuracy on the small EEGNet trunk is modest: 34.4% with fixed heads and 41.4% with the conditioned head, against a chance of 20.0%; fixed heads on frozen CBraMod features reach 70.7%. Balanced accuracy of each set-up: the mean over 100 people with its 95% interval. Chance is 20.0% for the sleep stage and 50.0% for the other two; the gate’s floors are 25.0% and 55.0%. | Set-up | SL-A · sleep stage, five stages | SL-E · next epoch’s stage differs | SL-F · first or second half of the night | | --- | --- | --- | --- | | EEGNet trained from scratch (E1) | | | | | Fixed heads — B-lin | 34.4% (95% interval 32.1%–37.0%) | 54.4% (95% interval 53.5%–55.4%) | 54.2% (95% interval 53.0%–55.6%) | | Shared hidden layer — B-sh | 42.9% (95% interval 40.1%–45.7%) | 58.9% (95% interval 57.5%–60.2%) | 55.8% (95% interval 54.2%–57.4%) | | Question-conditioned head — C1 | 41.4% (95% interval 38.6%–44.2%) | 57.5% (95% interval 56.0%–58.8%) | 55.6% (95% interval 54.2%–57.1%) | | Separate models — A | 35.3% (95% interval 32.9%–37.8%) | 53.2% (95% interval 52.2%–54.3%) | 52.2% (95% interval 51.0%–53.7%) | | Frozen CBraMod features (L1) | | | | | Fixed heads — B-lin-fz-sgd | 70.7% (95% interval 68.5%–72.7%) | 63.5% (95% interval 62.3%–64.7%) | 64.5% (95% interval 62.8%–66.1%) | | Shared hidden layer — B-sh-fz | 71.4% (95% interval 69.1%–73.6%) | 65.5% (95% interval 64.1%–66.7%) | 65.2% (95% interval 63.6%–66.9%) | | Question-conditioned head — C1-fz | 70.9% (95% interval 68.6%–73.0%) | 65.2% (95% interval 63.9%–66.5%) | 65.2% (95% interval 63.6%–66.9%) | ### A question that follows from another: does it need its own model? Awake or asleep (SL-B) follows from the sleep stage. A separate wake-or-sleep model, A(SL-B), left less error than reading the answer off the five-stage model A(SL-A): its remaining error was 0.684 times the read-out’s. The interval does not rule out a reduction of 20% or more, the margin set before the run, so “a derivable question needs no model of its own” is not supported. log R, the dedicated model’s remaining error (1 − AUROC) over the read-out’s, with its 95% interval; 100 people, three seeds. | Comparison | log R | R | AUROC, dedicated / read-out | Difference shown? | Margin, 20% | Gate | | --- | --- | --- | --- | --- | --- | --- | | P3 · Its own model − reading it off the five-stage model — A(SL-B) / A(SL-A) · SL-B | −0.380 (95% interval −0.583 to −0.203) | 0.684 | 0.941 / 0.914 | Dedicated model leaves less error | 20% margin not met | Gate passed | ### CBraMod adapted by LoRA (E2): secondary, one seed This arm ran after every other result of the run and two independent audits were known. Its design (35 LoRA fits, the seed, the recipe, the 30-second input as fifteen 2-second segments, the comparisons) and the condition for running it were fixed when the protocol was frozen, and nothing about it was chosen from results. It ran with temporary private checkpoints so that a fit could resume: during the run 39 were written for 33 fits, each replaced by its fit’s next one or deleted when its fit finished, so none outlived its fit, and 0 remain; a check of resuming, run before the fits, removed its own the same way. Its scoring ran the frozen code, and an independent audit recomputed all 9 entries. Paired differences in balanced accuracy, in pp, with 95% intervals; 100 people, one seed: no training-run variance in these intervals. | Comparison | Question | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | | S3-P1 · Conditioned head − fixed heads — C1 − B-lin | SL-A — sleep stage, five stages | +0.54 pp (95% interval −0.41 to +1.51 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P1 · Conditioned head − fixed heads — C1 − B-lin | SL-E — next epoch’s stage differs | +3.00 pp (95% interval +2.20 to +3.81 pp) | C1 higher | 2 pp margin not met | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P1 · Conditioned head − fixed heads — C1 − B-lin | SL-F — first or second half of the night | +0.80 pp (95% interval −0.19 to +1.79 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | SL-A — sleep stage, five stages | +0.15 pp (95% interval −0.15 to +0.44 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | SL-E — next epoch’s stage differs | +0.15 pp (95% interval −0.13 to +0.44 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P5 · Conditioned head − the same layer, not told the question — C1 − B-sh | SL-F — first or second half of the night | +0.03 pp (95% interval −0.23 to +0.29 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P2 · Fixed heads on one adapted CBraMod − separate models — B-lin − A(SL-A) | SL-A — sleep stage, five stages | +0.10 pp (95% interval −0.50 to +0.69 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P2 · Fixed heads on one adapted CBraMod − separate models — B-lin − A(SL-E) | SL-E — next epoch’s stage differs | −0.70 pp (95% interval −1.51 to +0.10 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | | S3-P2 · Fixed heads on one adapted CBraMod − separate models — B-lin − A(SL-F) | SL-F — first or second half of the night | −0.31 pp (95% interval −1.59 to +0.98 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the E2-sleep scorer, recomputed by its audit | With CBraMod adapted, the fixed heads were equivalent within ±2 pp of the conditioned head on the sleep stage and the half of the night; on the next change the conditioned head was +3.00 pp higher (+2.20 to +3.81 pp). Shared against separate models: equivalent within ±2 pp on all three. Gate arm, fixed heads at E2: SL-A 72.8% (70.9%–74.6%), SL-E 65.0% (63.8%–66.2%), SL-F 64.9% (63.3%–66.3%); all passed. What each set-up costs on sleep (EEGNet, E1). Step times were measured once, interleaved across set-ups on a shared GPU, and are indicative: in the cost rule a time difference counts only beyond 1.2×; the other columns are exact. | Set-up | Trainable parameters | Encoder passes per epoch | Head multiply-accumulates per epoch | Training steps per fit | Time per step | | --- | --- | --- | --- | --- | --- | | Fixed heads — B-lin | 34,905 (encoder 1,200 · heads 33,705 · conditioning 0) | 1 | 33,696 | 19,982 | 3.24 ms | | Shared hidden layer — B-sh | 241,593 (encoder 1,200 · heads 240,393 · conditioning 0) | 1 | 240,192 | 19,982 | 3.56 ms | | Question-conditioned head — C1 | 243,817 (encoder 1,200 · heads 240,393 · conditioning 2,224) | 1 | 240,384 | 19,982 | 4.04 ms | | Separate models — A | 37,305 (one EEGNet per question) | 3 | 33,696 | 59,916 | 2.60 ms | ### What the canary and the gate set aside The canary, on the first fold’s 80 training people, gave SL-A 36.8%, SL-B 59.9%, SL-C 70.7%, SL-D 82.1%, SL-E 56.5%, SL-F 58.4%, all below 85.0%, so none was demoted. The gate set aside the next change and the half of the night on EEGNet; every question passed on frozen CBraMod features and with CBraMod adapted. Secondary · crude replication · EESM19 ## The same comparison on EESM19 Every stored 30-second epoch of the full EESM19 release (first scorer) that passes the published core rules, at its natural stage mix: 52,311 epochs from 20 people. EEGNet from scratch, one seed, and no shared-hidden-layer arm, so the question code and the hidden layer are confounded here. A different protocol from the balanced scalp subset in the core matrix. SL-E and SL-F are largely predictable from the current stage; no stage-only read-out was computed here. Paired differences in balanced accuracy, in pp, with 95% intervals; 20 people, one seed. | Comparison | Question | The two arms’ means | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | --- | | S13-P1 · Conditioned head − fixed heads — C1 − B-lin | SL-A — sleep stage, five stages | 66.7% / 63.2% | +3.47 pp (95% interval +1.54 to +5.64 pp) | C1 higher | 2 pp margin not met | Gate passed — attached by the run | | S13-P1 · Conditioned head − fixed heads — C1 − B-lin | SL-E — next epoch’s stage differs | 56.8% / 53.4% | +3.38 pp (95% interval +2.27 to +4.36 pp) | C1 higher | 2 pp margin not met — not read: at floor | At floor: reported, not counted — attached by the run | | S13-P1 · Conditioned head − fixed heads — C1 − B-lin | SL-F — first or second half of the night | 63.8% / 61.8% | +2.09 pp (95% interval +0.38 to +3.74 pp) | C1 higher | 2 pp margin not met | Gate passed — attached by the run | | S13-P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(SL-A) | SL-A — sleep stage, five stages | 63.2% / 64.7% | −1.47 pp (95% interval −2.90 to −0.10 pp) | A(SL-A) higher | 2 pp margin not met | Gate passed — attached by the run | | S13-P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(SL-E) | SL-E — next epoch’s stage differs | 53.4% / 52.1% | +1.33 pp (95% interval +0.42 to +2.32 pp) | B-lin higher | B-lin non-inferior at 2 pp — not read: at floor | At floor: reported, not counted — attached by the run | | S13-P2 · Fixed heads on one small-CNN trunk − separate models — B-lin − A(SL-F) | SL-F — first or second half of the night | 61.8% / 62.5% | −0.72 pp (95% interval −2.10 to +0.69 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the run | On the sleep stage the direction matches BOAS: the conditioned head above the fixed heads (+3.47 pp, +1.54 to +5.64 pp), margin not met. The canary, on 16 training people, demoted nothing; the gate set the next change aside. Secondary · exploratory ## Secondary and exploratory results Further arms on motor imagery and sleep, one seed unless marked: every question trained at once, longer training, other encoders and other heads. Show the secondary results Codes: K-all, every question of the dataset trained together (MI-C and MI-D added on OpenBMI; SL-B, SL-C and SL-D on BOAS); primary K, the primary questions only; 2x, twice the training epochs; step-matched, A(MI-B) trained for the shared set-ups’ number of steps; B-MLP64-fz, a separate 64-unit hidden layer per question; H, a hypernetwork over learned question codes (exploratory: equivalent to linear heads by construction); C2-fz, query tokens over CBraMod’s token features. Each gate says who attached it: the run, the level’s gate, or the independent secondary audit, which attached the gates the frozen scoring left out. Motor imagery, OpenBMI | Entry | Level | Question | Contrast | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | --- | --- | | S2 — three seeds | E1 | MI-A — imagery or rest | B-sh − B-lin | +1.17 pp (95% interval −0.09 to +2.54 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the independent secondary audit | | S2 — three seeds | E1 | MI-B — which hand | B-sh − B-lin | −0.54 pp (95% interval −1.65 to +0.74 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S10 | E1 | MI-A — imagery or rest | B-lin(K-all) − B-lin(primary) | −2.32 pp (95% interval −3.94 to −0.75 pp) | B-lin(primary) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S10 | E1 | MI-B — which hand | B-lin(K-all) − B-lin(primary) | −1.84 pp (95% interval −3.26 to −0.41 pp) | B-lin(primary) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S10 | E1 | MI-A — imagery or rest | C1(K-all) − C1(primary) | −1.12 pp (95% interval −2.47 to +0.21 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the independent secondary audit | | S10 | E1 | MI-B — which hand | C1(K-all) − C1(primary) | −2.44 pp (95% interval −4.75 to −0.24 pp) | C1(primary) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S5 | E1 | MI-A — imagery or rest | C1(2x) − B-lin(2x) | +0.89 pp (95% interval −0.55 to +2.59 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the independent secondary audit | | S5 | E1 | MI-B — which hand | C1(2x) − B-lin(2x) | −0.24 pp (95% interval −1.43 to +0.98 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S6-P1 | E1 | MI-C — offline or online run | C1(K-all) − B-lin(K-all) | −0.15 pp (95% interval −2.86 to +2.50 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the independent secondary audit | | S6-P2 | E1 | MI-C — offline or online run | B-lin(K-all) − A(MI-C) | −2.75 pp (95% interval −5.43 to −0.22 pp) | A(MI-C) higher | 2 pp margin not met | Gate passed — attached by the run | | S6-P1 | E1 | MI-D — day 1 or day 2 | C1(K-all) − B-lin(K-all) | −0.69 pp (95% interval −4.17 to +2.64 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S6-P2 | E1 | MI-D — day 1 or day 2 | B-lin(K-all) − A(MI-D) | +0.49 pp (95% interval −3.23 to +3.81 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size — not read: at floor | At floor: reported, not counted — attached by the run | | S12 | E1 | MI-B — which hand | A(MI-B), step-matched − A(MI-B) | −0.03 pp (95% interval −1.63 to +1.44 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S12-P2 | E1 | MI-B — which hand | B-lin − A(MI-B), step-matched | −0.57 pp (95% interval −2.09 to +1.05 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the independent secondary audit | | S3-P1 | E2 | MI-A — imagery or rest | C1 − B-lin | −0.43 pp (95% interval −1.06 to +0.18 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S3-P2 | E2 | MI-A — imagery or rest | B-lin − A(MI-A) | −0.09 pp (95% interval −0.58 to +0.39 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S3-P5 | E2 | MI-A — imagery or rest | C1 − B-sh | −0.07 pp (95% interval −0.36 to +0.23 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S3-P1 | E2 | MI-B — which hand | C1 − B-lin | −1.07 pp (95% interval −2.08 to −0.08 pp) | B-lin higher | B-lin non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S3-P2 | E2 | MI-B — which hand | B-lin − A(MI-B) | +0.35 pp (95% interval −0.46 to +1.18 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S3-P5 | E2 | MI-B — which hand | C1 − B-sh | −0.16 pp (95% interval −0.71 to +0.42 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S7 | L1 · ST-EEGFormer Base, pooled | MI-A — imagery or rest | C1-fz − B-lin-fz-sgd | −1.11 pp (95% interval −1.89 to −0.32 pp) | B-lin-fz-sgd higher | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S7 | L1 · ST-EEGFormer Base, pooled | MI-B — which hand | C1-fz − B-lin-fz-sgd | −0.26 pp (95% interval −1.43 to +0.85 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S14 | L1 · CBraMod, pooled | MI-A — imagery or rest | B-MLP64-fz − B-sh-fz | −0.08 pp (95% interval −0.66 to +0.49 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | X1 | L1 · CBraMod, pooled | MI-A — imagery or rest | H − B-lin-fz-sgd | −1.78 pp (95% interval −2.30 to −1.24 pp) | B-lin-fz-sgd higher | B-lin-fz-sgd non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S14 | L1 · CBraMod, pooled | MI-B — which hand | B-MLP64-fz − B-sh-fz | +0.59 pp (95% interval −0.32 to +1.57 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | X1 | L1 · CBraMod, pooled | MI-B — which hand | H − B-lin-fz-sgd | −0.49 pp (95% interval −1.17 to +0.15 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S8 | L1 · CBraMod tokens | MI-A — imagery or rest | C2-fz − B-lin-fz-sgd | −0.95 pp (95% interval −2.27 to +0.45 pp) | No difference shown | B-lin-fz-sgd non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S8 | L1 · CBraMod tokens | MI-B — which hand | C2-fz − B-lin-fz-sgd | −5.31 pp (95% interval −7.09 to −3.69 pp) | B-lin-fz-sgd higher | B-lin-fz-sgd non-inferior at 2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S2 — three seeds — computed by the independent audit | L1 | MI-A — imagery or rest | B-sh-fz − B-lin-fz-sgd | +1.66 pp (95% interval +0.95 to +2.44 pp) | B-sh-fz higher | 2 pp margin not met | Gate passed — the level’s gate | | S2 — three seeds — computed by the independent audit | L1 | MI-B — which hand | B-sh-fz − B-lin-fz-sgd | +0.66 pp (95% interval −0.18 to +1.48 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — the level’s gate | **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. Sleep, BOAS | Entry | Level | Question | Contrast | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | --- | --- | | S2 — three seeds | E1 | SL-A — sleep stage, five stages | B-sh − B-lin | +8.43 pp (95% interval +5.83 to +11.12 pp) | B-sh higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S2 — three seeds | E1 | SL-E — next epoch’s stage differs | B-sh − B-lin | +4.43 pp (95% interval +3.10 to +5.71 pp) | B-sh higher | 2 pp margin not met — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S2 — three seeds | E1 | SL-F — first or second half of the night | B-sh − B-lin | +1.56 pp (95% interval −0.31 to +3.49 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S10 | E1 | SL-A — sleep stage, five stages | B-lin(K-all) − B-lin(primary) | −2.55 pp (95% interval −3.86 to −1.31 pp) | B-lin(primary) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S10 | E1 | SL-E — next epoch’s stage differs | B-lin(K-all) − B-lin(primary) | +0.97 pp (95% interval +0.23 to +1.71 pp) | B-lin(K-all) higher | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S10 | E1 | SL-F — first or second half of the night | B-lin(K-all) − B-lin(primary) | −0.61 pp (95% interval −1.37 to +0.16 pp) | No difference shown | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S10 | E1 | SL-A — sleep stage, five stages | C1(K-all) − C1(primary) | +4.14 pp (95% interval +2.76 to +5.49 pp) | C1(K-all) higher | C1(K-all) non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S10 | E1 | SL-E — next epoch’s stage differs | C1(K-all) − C1(primary) | +0.96 pp (95% interval +0.11 to +1.79 pp) | C1(K-all) higher | Equivalent within ±2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S10 | E1 | SL-F — first or second half of the night | C1(K-all) − C1(primary) | +1.67 pp (95% interval +0.55 to +2.87 pp) | C1(K-all) higher | C1(K-all) non-inferior at 2 pp — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S5 | E1 | SL-A — sleep stage, five stages | C1(2x) − B-lin(2x) | +1.71 pp (95% interval +0.12 to +3.25 pp) | C1(2x) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S5 | E1 | SL-E — next epoch’s stage differs | C1(2x) − B-lin(2x) | +2.13 pp (95% interval +1.27 to +2.93 pp) | C1(2x) higher | 2 pp margin not met — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S5 | E1 | SL-F — first or second half of the night | C1(2x) − B-lin(2x) | +2.78 pp (95% interval +1.72 to +3.94 pp) | C1(2x) higher | 2 pp margin not met — not read: at floor | At floor: reported, not counted — attached by the independent secondary audit | | S6-P1 | E1 | SL-C — REM or NREM | C1(K-all) − B-lin(K-all) | +5.50 pp (95% interval +4.04 to +7.04 pp) | C1(K-all) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S6-P2 | E1 | SL-C — REM or NREM | B-lin(K-all) − A(SL-C) | −9.59 pp (95% interval −11.77 to −7.58 pp) | A(SL-C) higher | 2 pp margin not met | Gate passed — attached by the run | | S6-P1 | E1 | SL-D — N3 or not | C1(K-all) − B-lin(K-all) | +9.79 pp (95% interval +6.69 to +12.95 pp) | C1(K-all) higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S6-P2 | E1 | SL-D — N3 or not | B-lin(K-all) − A(SL-D) | −2.13 pp (95% interval −4.50 to +0.07 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the run | | S11 — three seeds | E1 | SL-E — next epoch’s stage differs | B-lin − read-out from B-lin | +4.35 pp (95% interval +3.46 to +5.28 pp) | Dedicated head higher | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-E — next epoch’s stage differs | B-sh − read-out from B-sh | +8.26 pp (95% interval +6.91 to +9.52 pp) | Dedicated head higher | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-E — next epoch’s stage differs | C1 − read-out from C1 | +6.67 pp (95% interval +5.35 to +7.90 pp) | Dedicated head higher | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-E — next epoch’s stage differs | A(SL-E) − read-out from A(SL-A) | +3.09 pp (95% interval +2.02 to +4.14 pp) | Dedicated head higher | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-F — first or second half of the night | B-lin − read-out from B-lin | −0.35 pp (95% interval −1.47 to +0.87 pp) | No difference shown | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-F — first or second half of the night | B-sh − read-out from B-sh | −1.34 pp (95% interval −2.22 to −0.48 pp) | Read-out from the stage probabilities higher | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-F — first or second half of the night | C1 − read-out from C1 | −1.24 pp (95% interval −2.03 to −0.48 pp) | Read-out from the stage probabilities higher | No margin | No gate — descriptive | | S11 — three seeds | E1 | SL-F — first or second half of the night | A(SL-F) − read-out from A(SL-A) | −2.74 pp (95% interval −4.48 to −0.92 pp) | Read-out from the stage probabilities higher | No margin | No gate — descriptive | | S7 | L1 · REVE Large, pooled | SL-A — sleep stage, five stages | C1-fz − B-lin-fz-sgd | −0.86 pp (95% interval −1.60 to −0.11 pp) | B-lin-fz-sgd higher | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S7 | L1 · REVE Large, pooled | SL-E — next epoch’s stage differs | C1-fz − B-lin-fz-sgd | +2.30 pp (95% interval +1.36 to +3.20 pp) | C1-fz higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S7 | L1 · REVE Large, pooled | SL-F — first or second half of the night | C1-fz − B-lin-fz-sgd | +0.08 pp (95% interval −0.79 to +0.98 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | S14 | L1 · CBraMod, pooled | SL-A — sleep stage, five stages | B-MLP64-fz − B-sh-fz | −0.17 pp (95% interval −0.89 to +0.54 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | X1 | L1 · CBraMod, pooled | SL-A — sleep stage, five stages | H − B-lin-fz-sgd | −2.72 pp (95% interval −3.40 to −2.09 pp) | B-lin-fz-sgd higher | B-lin-fz-sgd non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S14 | L1 · CBraMod, pooled | SL-E — next epoch’s stage differs | B-MLP64-fz − B-sh-fz | −0.96 pp (95% interval −1.71 to −0.23 pp) | B-sh-fz higher | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | X1 | L1 · CBraMod, pooled | SL-E — next epoch’s stage differs | H − B-lin-fz-sgd | −1.94 pp (95% interval −2.74 to −1.14 pp) | B-lin-fz-sgd higher | B-lin-fz-sgd non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S14 | L1 · CBraMod, pooled | SL-F — first or second half of the night | B-MLP64-fz − B-sh-fz | −0.33 pp (95% interval −1.13 to +0.53 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — attached by the independent secondary audit | | X1 | L1 · CBraMod, pooled | SL-F — first or second half of the night | H − B-lin-fz-sgd | −1.50 pp (95% interval −2.29 to −0.73 pp) | B-lin-fz-sgd higher | B-lin-fz-sgd non-inferior at 2 pp | Gate passed — attached by the independent secondary audit | | S8 | L1 · CBraMod tokens | SL-A — sleep stage, five stages | C2-fz − B-lin-fz-sgd | +1.64 pp (95% interval +0.58 to +2.70 pp) | C2-fz higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S8 | L1 · CBraMod tokens | SL-E — next epoch’s stage differs | C2-fz − B-lin-fz-sgd | +2.93 pp (95% interval +2.02 to +3.83 pp) | C2-fz higher | 2 pp margin not met | Gate passed — attached by the independent secondary audit | | S8 | L1 · CBraMod tokens | SL-F — first or second half of the night | C2-fz − B-lin-fz-sgd | +1.05 pp (95% interval −0.03 to +2.11 pp) | No difference shown | 2 pp margin not met — inconclusive at this sample size | Gate passed — attached by the independent secondary audit | | S2 — three seeds — computed by the independent audit | L1 | SL-A — sleep stage, five stages | B-sh-fz − B-lin-fz-sgd | +0.72 pp (95% interval −0.13 to +1.64 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — the level’s gate | | S2 — three seeds — computed by the independent audit | L1 | SL-E — next epoch’s stage differs | B-sh-fz − B-lin-fz-sgd | +1.97 pp (95% interval +1.21 to +2.74 pp) | B-sh-fz higher | 2 pp margin not met | Gate passed — the level’s gate | | S2 — three seeds — computed by the independent audit | L1 | SL-F — first or second half of the night | B-sh-fz − B-lin-fz-sgd | +0.72 pp (95% interval −0.01 to +1.51 pp) | No difference shown | Equivalent within ±2 pp | Gate passed — the level’s gate | Questions that follow from the sleep stage (S4), log R | Entry | Contrast | log R | R | AUROC, dedicated / read-out | Difference shown? | Margin, 20% | Gate | | --- | --- | --- | --- | --- | --- | --- | --- | | S4-SL-C — 100 people | SL-C · A(SL-C) / A(SL-A) | −0.824 (95% interval −0.980 to −0.675) | 0.438 | 0.908 / 0.789 | Dedicated model leaves less error | 20% margin not met | Gate passed — attached by the run | | S4-SL-D — 71 people | SL-D · A(SL-D) / A(SL-A) | 0.126 (95% interval −0.002 to 0.251) | 1.134 | 0.975 / 0.978 | No difference shown | Read-out non-inferior at 20% | Gate passed — attached by the run | | S4-within-B-lin(K-all) — 100 people | SL-B · B-lin(K-all) / B-lin(K-all) | −0.525 (95% interval −0.665 to −0.395) | 0.592 | 0.946 / 0.908 | Dedicated model leaves less error | 20% margin not met | Gate passed — attached by the run | | S4-within-C1(K-all) — 100 people | SL-B · C1(K-all) / C1(K-all) | −0.088 (95% interval −0.144 to −0.031) | 0.916 | 0.946 / 0.941 | Dedicated model leaves less error | Equivalent within the 20% margin | Gate passed — attached by the run | Coherence (descriptive): the share of test epochs where a dedicated wake, REM or N3 head contradicts the same set-up’s five-stage head — B-lin(K-all): SL-B 17.2%, SL-C 21.0%, SL-D 1.0%; C1(K-all): SL-B 9.8%, SL-C 12.9%, SL-D 1.2%; A(SL-B) against A(SL-A): 16.5%. Methods & limits ## Limits - The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). - MI-A also separates a cue display and position in the run from imagery; MI-B is offline trials only; MI-C and MI-D are about the recording context, not brain states. - SL-B, SL-C and SL-D are deterministic functions of the 5-stage label. Results on them are about head design, not about new information in the signal. - SL-F is time of night, correlated with stage; SL-E concerns the next 30 s, not a prediction horizon in general. - With identifier-only conditioning, FiLM can differ from the same head without it only by per-question thresholds and signs of the shared hidden units. A null result does not show that language conditioning cannot help. - At E1 the shared encoder is a small CNN whose trunk has 1,200-2,096 parameters; P2 there is 'small-CNN trunk sharing', not foundation-model sharing; E2 on motor imagery ran at one seed, as a secondary arm. - SL-E and SL-F are largely predictable from the current stage; their results are shown beside a stage-only readout. - Questions answerable from window statistics or at floor or ceiling are reported but do not count for the route sentence. - Costs were measured on a shared GPU and are indicative; parameter counts, encoder passes and multiply-accumulates are exact. - Three datasets (EESM19 as a crude replication), one lab each, and one fixed recipe per encoder: this is not a ranking of encoders or of multi-task methods. - BOAS has natural stage prevalence but is one population and one PSG system; balanced accuracy across people is not a deployment error rate. - Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. - The OpenBMI numbers here use route 2's own 2-s windows, label-free high-pass and person-disjoint folds; they are not comparable with the site's OpenBMI cross-session calibration figures. - EESM19 full (S13) uses every stored epoch that passes the published core rules at its natural stage mix; it is a different protocol from the site's balanced sleep-scalp subset. - On EESM19 (S13) no stage-only readout was computed (S11 is defined on BOAS only), so its SL-E and SL-F entries must carry the sentence that both questions are largely predictable from the current stage. - E2 sleep ran after every other result was known; its design and run condition were fixed at the freeze, and nothing about it was chosen from results. ### Not run, and not reported - **Ying 2025 multi-night sleep (S9)**: deferred by the owner before the freeze; nothing was run - **E2 on motor imagery at three seeds**: not scheduled: E2 ran at one seed as a secondary arm, so the E1 sharing contrast keeps the label small-CNN trunk sharing - **The ridge reference row on frozen CBraMod features (G-1)**: Declared in the frozen draft, never implemented, and not computed afterwards: it would be new fits, and each choice it leaves open would be made after every result is known. No decision uses it. - **Arm-level results of the secondary arms (interval, AUROC, log loss, per-seed means)**: Not produced by the frozen stage 2; each secondary contrast carries the mean balanced accuracy of its two arms. - **P3 at E2**: Not declared for the E2-sleep block, so its five wake-or-sleep fits feed no contrast. - **Counts of people whose paired contrast is above or below zero**: Allowed by the BOAS review, not computed by the frozen stage 2, and not added afterwards. ### Checks Three independent audits re-implemented the definitions without the study code and recomputed the figures from the frozen predictions: the primary audit passed 10 of 10 checks over 1,673 values, the secondary 10 of 10 over 971, and the E2-sleep audit 25 of 25; none found a mismatch. ### The recordings ### OpenBMI motor imagery (Lee et al. 2019) Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. [Source ↗](https://doi.org/10.5524/100542) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) ### BOAS, the Bitbrain Open Access Sleep dataset Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. [Source ↗](https://openneuro.org/datasets/ds005555/versions/1.1.3) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) ### EESM19, full release (first scorer) Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. [Source ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) · [CC0-1.0 declared by upstream and mirror](https://creativecommons.org/publicdomain/zero/1.0/) ### Methods these results draw on (titles checked against arXiv on 7 October 2026) - [Yu & Yao, 2026 · Visual Jev: Accurate and Efficient Decisions from Shared Visual Context ↗](https://arxiv.org/abs/2609.25845) - [Dai et al., 2026 · Towards Unified Multi-task EEG Analysis with Low-Rank Adaptation ↗](https://arxiv.org/abs/2604.25131) - [Xiong et al., 2025 · EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models ↗](https://arxiv.org/abs/2508.17742) - [Perez et al., 2017 · FiLM: Visual Reasoning with a General Conditioning Layer ↗](https://arxiv.org/abs/1709.07871) - [Ha, Dai & Le, 2016 · HyperNetworks ↗](https://arxiv.org/abs/1609.09106) Data source: [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) · schema bci-report-shared-representation-update-v1. A longer write-up of the motor-imagery results is on Hugging Face: [Can one EEG encoder answer several questions? ↗](https://huggingface.co/blog/Twu31/one-eeg-encoder-several-questions) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Can one EEG model answer several questions about the same data as well as separate models, and at what cost?* https://bci.report/topics/shared-encoder/ Figures from release [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007) (2026-10-07). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/shared-encoder/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Can an EEG model answer questions asked in words? Questions in language · route 3 of the decision-research plan One small head over frozen EEG features, asked each question by a question number, a label template or a natural-language description: SSVEP on BETA, sleep on BOAS. Seen questions, rewordings and questions the EEG head was never trained on are reported separately. - [Reviewed aggregate JSON ↓](https://bci.report/data/questions-in-language-update.json) - Measured on: [BETA](https://bci.report/datasets/beta/); [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) · Methods: [CBraMod](https://bci.report/methods/cbramod/) ## Short answer Only for questions it was trained on, and not on every dataset. Asked a seen question with a label template or a description instead of a question number, a small head on frozen EEG features was equivalent within the 2 pp margin on sleep (BOAS) but lost accuracy on SSVEP (BETA, **70** people): **−5.96 pp** with templates on a plain spectrum (**−6.84** to **−5.08 pp**), the margin not met. Rewording a seen question cost accuracy too, except a new sentence frame on sleep. Questions the EEG head was never trained on were not answered usefully: every unseen sleep question did worse than adding up the head’s own stage answers, and on unseen flicker frequencies the head reached **28.7%** on a plain spectrum where CCA, which needs no training, reached **80.9%** (chance **12.5%**). Explicit wordings such as “N1 or N2 sleep” used the link between words and stages; the everyday names “asleep” and “light sleep” ranked windows backwards, and in a boundary probe negated questions were answered as if they asked for what they negate. How the comparison works ## One head, three ways to ask The same small head answers every question; only the vector that tells it which question changes. Every test is on people the head never trained on. - **ID · question number**: a one-hot code for each question the head was trained on. It cannot ask anything new. - **TPL · label template**: a sentence frame filled with the label’s name, turned into a vector by a frozen text encoder. - **DESC · description**: a sentence that describes the state without naming it, turned into a vector the same way. - **SHUF · shuffled templates**: the template vectors attached to the wrong questions during training, by derangements declared in advance: a control for whether the head uses the link between a wording and its question. - **NUM · numeric code**: for SSVEP, the frequency coded as numbers: a control without text. - **B-sh · one output per question**: the same hidden layer with a fixed output per seen question, as in route 2; it cannot ask anything new either. - **CCA · no training**: for SSVEP, canonical correlation analysis, which needs no training and can name any frequency: the reference an unseen frequency is held to. A training frame for sleep, filled with a stage’s name: The scorer marked this 30-second epoch as N2 sleep. SSVEP on BETA: a plain spectrum (L0) and frozen CBraMod features (L1). Sleep on BOAS: frozen CBraMod features (L1). Three training seeds each. The primary text encoder is bge-small-en-v1.5, frozen. The head is the same for every arm: standardised features, a small hidden layer adjusted by the question vector (FiLM), one output. Three kinds of result, never pooled into one zero-shot number: a seen question in the wording the head trained with; the same question reworded; and a question the head never trained on. “Unseen” always means unseen by the EEG head, not by the text encoder: the frozen encoder has read the words. Each comparison is the language arm minus the reference, in percentage points (pp), with a 95% interval from resampling people, then training seeds, then held-out wordings. It gets two readings, both fixed before any fit. A difference is shown when the interval excludes zero. Against a margin of ±2 pp, it is equivalent when the whole interval lies inside, non-inferior when the bound against the language arm stays within the margin, and otherwise the margin is not met, which is not the same as a loss of 2 pp or more. For an unseen sleep question the comparison is R, the head’s remaining error (1 − AUROC) over that of the read-off from its own stage answers, with a margin of R < 1.25. The check of whether the head uses the link between wording and stage, the template arm against the shuffled templates, is a difference of AUROC. Among the primary comparisons, those with the numeric code, the neighbour average and the shuffled templates are read for a difference only; the secondary tables print both readings as the run computed them. There are 35 pre-declared primary comparisons and no multiplicity correction, so at 95% about one in twenty comparisons with no true difference may show one by chance. Accuracy is averaged per person, then over people, to compare set-ups. “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. The decision-research plan started from a vision paper, [Yu & Yao, 2026 · Visual Jev: Accurate and Efficient Decisions from Shared Visual Context ↗](https://arxiv.org/abs/2609.25845), which applies that idea to images. All three routes: [Jev-style questions on EEG: the evidence →](https://bci.report/jev-style/) Primary · seen questions · BETA and BOAS ## Seen questions: asked in words, and reworded On BETA, 70 people look at flickering targets; in each rotation the head trains on most of the frequencies and is asked about those (the 32-way task). On BOAS, 100 people, each half-minute epoch is scored into one of five stages at their natural mix, against the human consensus. **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. ### In the training wording (P1) **−5.96 pp** TPL − ID on SSVEP, a plain spectrum, BETA, three seeds ### Sleep: no measurable cost. SSVEP: a cost. On sleep, a template or a description instead of a question number was equivalent within ±2 pp: −0.05 pp (−0.70 to +0.61 pp) and +0.02 pp (−0.59 to +0.62 pp). On SSVEP it cost accuracy at both levels, and the margin was not met: templates −5.96 pp on a plain spectrum and −3.49 pp on frozen CBraMod features; descriptions −5.46 pp and −3.37 pp. Paired differences in balanced accuracy, language arm minus question number, in pp, with 95% intervals; BETA 70 people, BOAS 100 people, three seeds each. | Comparison | Question | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | | SSVEP · BETA · a plain spectrum (L0) | | | | | | | P1 · Label template − question number — TPL − ID | seen, 32-way | −5.96 pp (95% interval −6.84 to −5.08 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 32-way | −5.46 pp (95% interval −6.26 to −4.66 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | SSVEP · BETA · frozen CBraMod features (L1) | | | | | | | P1 · Label template − question number — TPL − ID | seen, 32-way | −3.49 pp (95% interval −4.10 to −2.93 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 32-way | −3.37 pp (95% interval −3.98 to −2.80 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | Sleep · BOAS · frozen CBraMod features (L1) | | | | | | | P1 · Label template − question number — TPL − ID | seen, 5-way | −0.05 pp (95% interval −0.70 to +0.61 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 5-way | +0.02 pp (95% interval −0.59 to +0.62 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | The seen SSVEP task is hard for every head on these features: with question numbers the head reached 32.1% (28.8%–35.2%) on a plain spectrum and 24.2% on frozen CBraMod features, against a chance of 3.1%; CCA reached 65.5%. On sleep every arm reached about the same balanced accuracy, 71.5% with question numbers against a chance of 20.0%. The paired differences above are the comparison; in the levels below, one arm is above another only where their 95% intervals do not overlap. BETA, seen 32-way balanced accuracy (chance 3.1%): the mean over 70 people with its 95% interval, and the chance-corrected value, (level − chance) / (1 − chance). Three seeds, CCA none. | Arm | A plain spectrum (L0) | chance-corrected | Frozen CBraMod features (L1) | chance-corrected | | --- | --- | --- | --- | --- | | ID — question number | 32.1% (95% interval 28.8%–35.2%) | 29.9% | 24.2% (95% interval 21.6%–26.9%) | 21.8% | | TPL — label template | 26.1% (95% interval 23.5%–28.6%) | 23.7% | 20.7% (95% interval 18.5%–23.1%) | 18.2% | | DESC — description | 26.6% (95% interval 23.9%–29.2%) | 24.2% | 20.9% (95% interval 18.7%–23.2%) | 18.3% | | SHUF — shuffled templates | 25.1% (95% interval 22.5%–27.6%) | 22.7% | 20.2% (95% interval 18.1%–22.3%) | 17.6% | | NUM — numeric code | 23.7% (95% interval 21.2%–26.1%) | 21.2% | 13.9% (95% interval 12.5%–15.2%) | 11.1% | | B-sh — one output per question | 32.2% (95% interval 28.9%–35.3%) | 30.0% | 27.0% (95% interval 24.2%–30.1%) | 24.7% | | CCA — no training | 65.5% (95% interval 59.7%–71.1%) | 64.3% | 65.5% (95% interval 59.7%–71.1%) | 64.3% | BOAS, seen five-stage balanced accuracy (chance 20.0%), 100 people, three seeds, with the chance-corrected value and the log loss (mean binary cross-entropy, in nats). The shuffled-template arm’s seen score is read through its own derangement, and it has no log loss. | Arm | Balanced accuracy | chance-corrected | Log loss | | --- | --- | --- | --- | | ID — question number | 71.5% (95% interval 69.1%–73.5%) | 64.3% | 0.267 (95% interval 0.242–0.296) | | TPL — label template | 71.4% (95% interval 69.1%–73.4%) | 64.3% | 0.276 (95% interval 0.249–0.307) | | DESC — description | 71.5% (95% interval 69.1%–73.5%) | 64.3% | 0.278 (95% interval 0.251–0.310) | | SHUF — shuffled templates | 71.5% (95% interval 69.2%–73.5%) | 64.4% | — | | B-sh — one output per question | 71.4% (95% interval 69.1%–73.4%) | 64.3% | 0.259 (95% interval 0.236–0.287) | BOAS, AUROC of each seen stage, with 95% intervals; three seeds. N3 pools the 71 people with at least five N3 epochs; every other stage pools 100. | Arm | W — 100 people | N1 — 100 people | N2 — 100 people | N3 — 71 people | REM — 100 people | | --- | --- | --- | --- | --- | --- | | ID — question number | 0.982 (0.976–0.987) | 0.876 (0.859–0.890) | 0.922 (0.904–0.937) | 0.978 (0.971–0.983) | 0.957 (0.945–0.967) | | TPL — label template | 0.980 (0.972–0.987) | 0.872 (0.855–0.887) | 0.924 (0.909–0.937) | 0.979 (0.974–0.983) | 0.956 (0.944–0.967) | | DESC — description | 0.981 (0.974–0.987) | 0.869 (0.851–0.884) | 0.923 (0.907–0.937) | 0.979 (0.974–0.983) | 0.956 (0.944–0.967) | ### Reworded (P2) The same seen questions in wordings the head never trained with: a new sentence frame (a1, held-out frames resampled), a synonym of the label (a2: three per question, fixed, so these entries hold only for these three synonyms), or a new description (a3, held-out descriptions resampled). Paired differences in balanced accuracy, reworded minus the training wording, in pp, with 95% intervals; BETA 70 people, BOAS 100 people, three seeds each. | Comparison | Question | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | | SSVEP · BETA · a plain spectrum (L0) | | | | | | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −4.01 pp (95% interval −6.53 to −2.12 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −1.84 pp (95% interval −2.10 to −1.57 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −4.13 pp (95% interval −5.58 to −3.11 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | SSVEP · BETA · frozen CBraMod features (L1) | | | | | | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −2.72 pp (95% interval −4.43 to −1.43 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −1.33 pp (95% interval −1.53 to −1.14 pp) | TPL(train) higher | Equivalent within ±2 pp | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −2.95 pp (95% interval −3.81 to −2.26 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | Sleep · BOAS · frozen CBraMod features (L1) | | | | | | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −0.46 pp (95% interval −1.33 to +0.16 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −12.60 pp (95% interval −13.74 to −11.43 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −20.87 pp (95% interval −29.49 to −13.71 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | On sleep a new sentence frame cost nothing measurable: equivalent within ±2 pp (−0.46 pp, −1.33 to +0.16 pp). The three synonyms per stage scored −12.60 pp (−13.74 to −11.43 pp) and new descriptions −20.87 pp (−29.49 to −13.71 pp), a wide interval, because descriptions differ a lot from one another. On BETA every rewording scored lower, from −1.33 pp to −4.13 pp; only the synonyms on frozen CBraMod features stayed within the margin. Each synonym on its own against the training wording, in pp with 95% intervals; BETA 70 people, BOAS 100 people. On SSVEP the third synonym spells the number out. | Synonym | SSVEP · BETA · a plain spectrum (L0) | SSVEP · BETA · frozen CBraMod features (L1) | Sleep · BOAS · frozen CBraMod features (L1) | | --- | --- | --- | --- | | Synonym 1 | −1.45 pp (95% interval −1.67 to −1.22 pp) | −1.09 pp (95% interval −1.27 to −0.90 pp) | −23.53 pp (95% interval −25.29 to −21.64 pp) | | Synonym 2 | −0.87 pp (95% interval −1.04 to −0.70 pp) | −0.84 pp (95% interval −1.00 to −0.68 pp) | −10.85 pp (95% interval −12.59 to −9.17 pp) | | Synonym 3 | −3.19 pp (95% interval −3.69 to −2.69 pp) | −2.06 pp (95% interval −2.44 to −1.73 pp) | −3.42 pp (95% interval −4.49 to −2.41 pp) | Answer flip rate: how often the head’s answer to the same window changes when only the wording changes; BETA 70 people, BOAS 100 people. On SSVEP even another training seed of the question-number head changes most answers, because the seen task is near its floor, so flip rates there mostly show uncertainty. | | New frame | Synonym | Between training frames | New description | Between training descriptions | Between seeds of ID | | --- | --- | --- | --- | --- | --- | --- | | BETA · L0 | 39.7% | 29.3% | 19.4% | 41.2% | 19.3% | 56.8% | | BETA · L1 | 40.7% | 32.0% | 24.3% | 42.7% | 23.3% | 62.2% | | BOAS · L1 | 6.4% | 20.0% | 3.3% | 45.4% | 3.1% | 7.4% | Primary · unseen sleep questions · BOAS ## Sleep questions the EEG head was never trained on Three questions, each a combination of the stages the head trained on: asleep (N1, N2, N3 or REM), N1, N2 or N3 sleep, and light sleep (N1 or N2). Each was asked with templates, using an implicit name or an explicit one, and with new descriptions; 100 people, three seeds. Unseen by the EEG head, not by the text encoder. Because each question is a combination, the head’s own stage answers can be added up into an answer, the read-off, which is what each wording is held to. **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. The names the template arm was asked with, as written before any score. | Question | Implicit name | Explicit name | | --- | --- | --- | | Asleep | asleep | N1, N2, N3 or REM sleep | | N1, N2 or N3 sleep | none | N1, N2 or N3 sleep | | Light sleep | light sleep | N1 or N2 sleep | ### Can the head answer them? (P3) **33.80** R for “asleep” with its implicit name: the head’s remaining error over the read-off’s ### Not answered: the read-off did better in every wording In all eight conditions the interval lay above R = 1, in the read-off’s favour, and none came within the 1.25 margin: R ran from 2.42 (2.07–2.85; N1, N2 or N3 sleep, explicit name) to 33.80 (23.09–52.82; asleep, implicit name). log R and R, the head’s remaining error (1 − AUROC) over the read-off’s, with 95% intervals; 100 people, three seeds. A positive log R, or R above 1, is against the language arm. | Comparison | Question | log R, with R | Difference shown? | Margin, R < 1.25 | Gate | | --- | --- | --- | --- | --- | --- | | P3 · Label template head against its own read-off — TPL / read-off | asleep · implicit name | +3.520 (95% interval +3.139 to +3.967) (R 33.80 (23.09–52.82)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | asleep · explicit name | +1.951 (95% interval +1.623 to +2.357) (R 7.04 (5.07–10.56)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | N1, N2 or N3 sleep · explicit name | +0.883 (95% interval +0.726 to +1.047) (R 2.42 (2.07–2.85)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · implicit name | +2.267 (95% interval +2.072 to +2.466) (R 9.65 (7.94–11.77)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · explicit name | +1.519 (95% interval +1.343 to +1.698) (R 4.57 (3.83–5.46)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | asleep · new descriptions — 10 held-out wordings, resampled | +3.444 (95% interval +2.913 to +3.956) (R 31.30 (18.41–52.26)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | N1, N2 or N3 sleep · new descriptions — 10 held-out wordings, resampled | +1.299 (95% interval +0.793 to +1.744) (R 3.67 (2.21–5.72)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | light sleep · new descriptions — 10 held-out wordings, resampled | +1.417 (95% interval +0.903 to +1.856) (R 4.12 (2.47–6.40)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | AUROC of the template head on each unseen question, and of the read-off from its own stage answers (prior-corrected), with 95% intervals; 100 people, three seeds. An AUROC below 0.5 ranks windows in reverse. | Question | Wording | Template head | Read-off from its stage answers | | --- | --- | --- | --- | | Asleep | implicit name | 0.317 (95% interval 0.294–0.341) | 0.980 (95% interval 0.971–0.987) | | Asleep | explicit name | 0.858 (95% interval 0.839–0.875) | 0.980 (95% interval 0.971–0.987) | | N1, N2 or N3 sleep | explicit name | 0.871 (95% interval 0.854–0.887) | 0.947 (95% interval 0.934–0.958) | | Light sleep | implicit name | 0.244 (95% interval 0.222–0.267) | 0.922 (95% interval 0.907–0.935) | | Light sleep | explicit name | 0.642 (95% interval 0.616–0.667) | 0.922 (95% interval 0.907–0.935) | The implicit names ranked windows in reverse: AUROC 0.317 for “asleep” and 0.244 for “light sleep”, below 0.5, so windows that really were asleep, or in light sleep, scored lower on these questions than the others. The explicit names did better (0.858 and 0.642) but stayed short of the read-off (0.980 and 0.922). Descriptions have no alignment control: the shuffled control is built from template vectors. Held-out sentence frames (S13, in the secondary panel) give the same result for all five template cells. ### Is the link between wording and stage used? (P5) The shuffled-template arm attaches the same template vectors to the wrong stages during training; the unseen questions keep their true vectors. If the head uses the link between a wording and its stage, the shuffled arm should do worse on them than the template arm. Template arm minus shuffled templates: AUROC on the unseen questions, averaged over the cells of each kind of name, with 95% intervals; 100 people, three seeds. | Comparison | Question | Difference, AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, implicit names | −0.378 (95% interval −0.501 to −0.253) | SHUF higher | Difference only: no margin | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, explicit names | +0.629 (95% interval +0.603 to +0.654) | TPL higher | Difference only: no margin | Headroom gate passed | AUROC on each unseen question, template arm and shuffled templates, with 95% intervals; 100 people, three seeds. | Question | Wording | Template arm | Shuffled templates | | --- | --- | --- | --- | | Asleep | implicit name | 0.317 (95% interval 0.294–0.341) | 0.676 (95% interval 0.529–0.812) | | Asleep | explicit name | 0.858 (95% interval 0.839–0.875) | 0.137 (95% interval 0.109–0.168) | | N1, N2 or N3 sleep | explicit name | 0.871 (95% interval 0.854–0.887) | 0.171 (95% interval 0.145–0.201) | | Light sleep | implicit name | 0.244 (95% interval 0.222–0.267) | 0.641 (95% interval 0.537–0.738) | | Light sleep | explicit name | 0.642 (95% interval 0.616–0.667) | 0.176 (95% interval 0.155–0.199) | For explicit names, yes: attaching the templates to the wrong stages lowered the unseen AUROC, TPL − SHUF +0.629 (+0.603 to +0.654). For the implicit names it was not shown: the shuffled templates scored higher, −0.378 (−0.501 to −0.253). Primary · unseen flicker frequencies · BETA ## Flicker frequencies the EEG head was never trained on (P4) In each of five rotations, eight of BETA’s frequencies are held out: the head trains on the rest and is asked about the eight by stating the frequency in a template. 70 people, three seeds, at both levels. Unseen by the EEG head, not by the text encoder: every token of a held-out frequency also occurs in that rotation’s training texts, so only the combination is new. **−52.21 pp** TPL − CCA, 8-way among unseen frequencies, a plain spectrum, BETA ### Far below CCA In the 8-way among unseen frequencies the head reached 28.7% (26.9%–30.5%) on a plain spectrum and 19.0% on frozen CBraMod features; CCA reached 80.9%, and chance is 12.5%. The head also did worse than averaging its own answers for the two neighbouring seen frequencies, and worse than a numeric code of the frequency, at both levels. Paired differences in accuracy, template head minus the reference, in pp, with 95% intervals; 70 people, three seeds. | Comparison | Question | Difference, pp | Difference shown? | Margin, ±2 pp | Gate | | --- | --- | --- | --- | --- | --- | | SSVEP · BETA · a plain spectrum (L0) | | | | | | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, 8-way | −52.21 pp (95% interval −55.13 to −48.99 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − average of the two neighbouring seen frequencies — TPL − NN(TPL) | unseen frequency, 8-way | −10.98 pp (95% interval −12.24 to −9.67 pp) | NN(TPL) higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, 8-way | −12.77 pp (95% interval −14.52 to −10.99 pp) | NUM higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, neighbour 3-way | −62.36 pp (95% interval −65.87 to −58.69 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, neighbour 3-way | −1.24 pp (95% interval −3.17 to +0.85 pp) | No difference shown | Difference only: no margin | Headroom gate passed | | SSVEP · BETA · frozen CBraMod features (L1) | | | | | | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, 8-way | −61.94 pp (95% interval −65.65 to −58.06 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − average of the two neighbouring seen frequencies — TPL − NN(TPL) | unseen frequency, 8-way | −1.16 pp (95% interval −1.97 to −0.37 pp) | NN(TPL) higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, 8-way | −2.46 pp (95% interval −3.23 to −1.69 pp) | NUM higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, neighbour 3-way | −54.67 pp (95% interval −57.77 to −51.47 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, neighbour 3-way | +15.61 pp (95% interval +14.25 to +16.93 pp) | TPL higher | Difference only: no margin | Headroom gate passed | In the neighbour 3-way the unseen frequency competes with the trained detectors of its two neighbours, which favours the seen questions. On a plain spectrum the template head fell below the mean chance of that test, 19.3% against 34.2%, and showed no difference against the numeric code. On frozen CBraMod features it beat the numeric code by +15.61 pp, but those features keep the stimulus phase and neighbouring targets differ in phase, so that comparison tests frequency and phase together. The plain spectrum, which carries no phase, is the clean frequency test. Accuracy on the unseen frequencies: u8, the 8-way among the held-out frequencies (chance 12.5%); nn8, the read-off that averages the answers for the two neighbouring seen frequencies; u3, the neighbour 3-way (mean chance 34.2%, a 2-way at the two grid edges). 70 people; three seeds, CCA none. | Arm | Test | A plain spectrum (L0) | chance-corrected | Frozen CBraMod features (L1) | chance-corrected | | --- | --- | --- | --- | --- | --- | | TPL — label template | u8 | 28.7% (95% interval 26.9%–30.5%) | 18.5% | 19.0% (95% interval 18.0%–19.9%) | 7.4% | | DESC — description | u8 | 28.6% (95% interval 26.8%–30.4%) | 18.4% | 18.6% (95% interval 17.6%–19.6%) | 7.0% | | SHUF — shuffled templates | u8 | 9.2% (95% interval 8.6%–9.7%) | -3.8% | 11.0% (95% interval 10.5%–11.5%) | -1.7% | | NUM — numeric code | u8 | 41.5% (95% interval 38.2%–44.7%) | 33.1% | 21.4% (95% interval 20.2%–22.6%) | 10.2% | | CCA — no training | u8 | 80.9% (95% interval 76.6%–84.8%) | 78.2% | 80.9% (95% interval 76.6%–84.8%) | 78.2% | | TPL — label template | nn8 | 39.7% (95% interval 36.8%–42.5%) | 31.0% | 20.1% (95% interval 19.0%–21.3%) | 8.7% | | ID — question number | nn8 | 44.3% (95% interval 40.8%–47.6%) | 36.4% | 18.3% (95% interval 17.3%–19.3%) | 6.6% | | B-sh — one output per question | nn8 | 44.3% (95% interval 40.9%–47.5%) | 36.3% | 17.6% (95% interval 16.6%–18.8%) | 5.8% | | TPL — label template | u3 | 19.3% (95% interval 18.2%–20.5%) | -22.5% | 27.0% (95% interval 26.0%–28.0%) | -10.8% | | NUM — numeric code | u3 | 20.6% (95% interval 19.3%–21.8%) | -20.6% | 11.4% (95% interval 10.7%–12.1%) | -34.5% | | CCA — no training | u3 | 81.7% (95% interval 78.5%–84.6%) | 72.2% | 81.7% (95% interval 78.5%–84.6%) | 72.2% | Generalisation cost: accuracy on the unseen frequencies minus accuracy on matched seen ones, in pp with 95% intervals; 70 people. | Arm and test | A plain spectrum (L0) | Frozen CBraMod features (L1) | | --- | --- | --- | | ID, 8-way (neighbour read-off) | −14.26 pp (95% interval −15.50 to −13.06 pp) | −27.80 pp (95% interval −30.61 to −25.03 pp) | | TPL, 8-way | −24.44 pp (95% interval −26.44 to −22.26 pp) | −24.27 pp (95% interval −26.72 to −21.85 pp) | | TPL, 3-way | −27.83 pp (95% interval −30.18 to −25.37 pp) | −29.67 pp (95% interval −31.76 to −27.64 pp) | | DESC, 8-way | −24.79 pp (95% interval −26.88 to −22.55 pp) | −24.81 pp (95% interval −27.31 to −22.39 pp) | | DESC, 3-way | −28.95 pp (95% interval −31.33 to −26.46 pp) | −29.74 pp (95% interval −31.76 to −27.76 pp) | | SHUF, 8-way | −40.63 pp (95% interval −44.16 to −36.98 pp) | −30.09 pp (95% interval −33.02 to −27.27 pp) | | SHUF, 3-way | −34.96 pp (95% interval −37.50 to −32.43 pp) | −32.52 pp (95% interval −35.18 to −29.84 pp) | | NUM, 8-way | −12.30 pp (95% interval −13.42 to −11.17 pp) | −16.16 pp (95% interval −17.88 to −14.47 pp) | | NUM, 3-way | −7.27 pp (95% interval −8.49 to −6.02 pp) | −7.63 pp (95% interval −8.74 to −6.53 pp) | | B-sh, 8-way (neighbour read-off) | −12.81 pp (95% interval −14.09 to −11.56 pp) | −31.19 pp (95% interval −34.09 to −28.32 pp) | | CCA, 8-way | 0.00 pp (95% interval 0.00 to 0.00 pp) | 0.00 pp (95% interval 0.00 to 0.00 pp) | | CCA, 3-way | +0.41 pp (95% interval +0.26 to +0.56 pp) | +0.41 pp (95% interval +0.26 to +0.56 pp) | TPL − CCA in the 8-way, rotation by rotation, in pp with 95% intervals (secondary, recomputed by an independent audit); 70 people, three seeds. The head is below CCA in every rotation. | Rotation | A plain spectrum (L0) | Frozen CBraMod features (L1) | | --- | --- | --- | | Rotation 0 | −49.88 pp (95% interval −53.16 to −46.16 pp) | −63.00 pp (95% interval −67.23 to −58.54 pp) | | Rotation 1 | −52.96 pp (95% interval −57.02 to −48.99 pp) | −61.76 pp (95% interval −65.81 to −57.74 pp) | | Rotation 2 | −48.74 pp (95% interval −51.97 to −45.42 pp) | −59.74 pp (95% interval −63.75 to −55.51 pp) | | Rotation 3 | −57.34 pp (95% interval −60.84 to −53.73 pp) | −62.75 pp (95% interval −66.73 to −58.42 pp) | | Rotation 4 | −52.14 pp (95% interval −55.80 to −48.20 pp) | −62.47 pp (95% interval −66.53 to −58.23 pp) | Per-frequency and interior-only breakdowns are not published: they were computed after the audits, which do not cover them. Fourteen follow-up experiments, thirteen of them on the same BETA cohort, tested how the question is asked, the head, phase, the matching computation, a learned correction to CCA, the cross-spectral matrix, how CCA’s statistics are pooled and weighted, and filter-bank CCA with shorter windows and fewer electrodes; five ran on a second dataset, one of them only there: [what lets a decoder answer an SSVEP frequency it never trained on →](https://bci.report/topics/new-ssvep-frequencies/) Before any result ## The pre-run checks, and how one was changed Every check had to pass before any result was read. A training-label permutation canary refits the heads on shuffled training labels and checks that test scores sit at chance: 17 configurations on the first fold, 38 intervals at 99.9%. Revision 3 failed: 1 of 38 intervals excluded chance (sleep, frozen CBraMod features, the shuffled-template arm, unseen “light sleep” with its explicit name). Its labels had been permuted within each training person. The run stopped. Revision 4, approved by the owner before any new fit, permuted labels across all training windows. It failed too: 3 of 38 intervals, all on BOAS, one of them below chance; every SSVEP interval included chance both times. Revision 4 said “no further revision”. The owner overrode that after seeing how it failed, and approved revision 5 before any new fit. Revision 4’s per-person centring diagnostic was not run: with labels permuted across all windows it no longer tests anything. Revision 5 passed: 20 heads per configuration (340 fits), each metric compared with its exact per-person chance, and a two-level bootstrap over heads, then people; 0 of 38 intervals excluded zero. Both failed records are kept. Direction counts: of 760 single-head intervals in revision 5, 125 excluded the exact chance, 66 above and 59 below, and 18 cells failed in both directions. BOAS cells are published as pass or fail and counts only. Why they failed, the reading revision 5 was built to test: a head trained on shuffled labels is still a random function of sleep features that are strongly organised by stage, so on new people its scores lean one way for everyone, up or down. A person bootstrap measures the spread between people, not between heads, so its 99.9% interval was too narrow for this question. The failures fit that reading: all were on BOAS, and one was below chance, which a leak cannot produce. It is not a proof that no leak exists: person-disjoint folds, no held-out label or wording in training and training-only scaling are asserted in the code, and both conformance audits check for leakage directly. Decision S0b-6, approved by the owner: the SSVEP token rule says every token of a held-out frequency occurs in that rotation’s training texts. It stops the run for bge-small-en-v1.5 and all-MiniLM-L6-v2, and both passed in all 5 rotations. It cannot hold for multilingual-e5-small, whose tokenizer keeps one-decimal number pieces as single tokens while the held-out number itself is barred from training texts; for that encoder the rule was reported instead of stopping the run, and every Chinese-wording unseen-frequency entry says the frequency is a new token. Engineering amendment EA-1, before the first fit: in the contiguous extrapolation band (S8) the neighbour average is not defined and the neighbour 3-way is relabelled, because its competitors are themselves held-out questions; failed fits are marked per cell, and none failed; the secondary scoring can resume after an interruption. No question, partition, wording, arm, metric, margin or gate changed. The declared numeracy prediction was not activated. It said: if the primary text encoder orders frequencies poorly (Spearman ρ below 0.3 between frequency distance and embedding distance, for every training frame), the 8-way TPL − NUM flag will not read “TPL higher”. Its premise was false: ρ was 0.593, 0.563 and 0.526 for the three training frames of bge-small-en-v1.5, over 780 pairs. The flag reads “NUM higher” at both levels anyway; with the premise false, that is a description, not a confirmed prediction. ### The gates - Headroom: the question-number head’s seen balanced accuracy must lie above chance + 5 pp and below 95%. It passed on all three primary blocks. - Triviality canary: a ridge classifier on per-channel mean and log-variance must not answer the seen task by itself. It passed on BETA (on the first fold’s 60 training people) and on BOAS and EESM19 (carried from route 2); none is declared for motor imagery, whose gate reads “not run” and excludes nothing. - Read-off floor: an unseen sleep cell counts only if the read-off’s AUROC interval lies above 0.55. 8 of 8 primary and 34 of 34 secondary cells passed. - Shuffled-EEG scorer check: each primary aggregate, recomputed with the test features shuffled within each person, must sit within three Monte-Carlo standard errors of chance. 51 of 51 cells did (26 SSVEP, 25 BOAS). - Fit check, on the first fold’s training rows: every conditioned SSVEP arm trained more than 5 pp below the one-output-per-question head, so the declared fallback of 60 epochs applies to every SSVEP arm; sleep, EESM19 and motor imagery stay at 30. No fit failed. Four independent audits passed. They re-implement the definitions without importing the study’s scoring code: the conformance audit of the primary run passed 106 checks with 0 failed; the numerical audit of the primary made 2,270 comparisons with 0 mismatches; the conformance audit of the secondary run passed 231 checks with 0 failed; the numerical audit of the secondary made 18,619 comparisons, 29 of them outside the tolerance, every one explained. Secondary · beyond the primary comparisons ## Secondary results Further datasets, features, a second text encoder, Chinese wordings and probes. One seed unless marked: a single seed carries no training-run variance. Every EESM19 and single-seed entry was declared likely inconclusive before the run; none reached the “inconclusive” flags. Show the secondary results ### S1 · EESM19 sleep, frozen CBraMod features · 20 people, three seeds · declared likely inconclusive | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 5-way | −0.04 pp (95% interval −0.47 to +0.39 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 5-way | −0.03 pp (95% interval −0.43 to +0.40 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −0.11 pp (95% interval −0.56 to +0.20 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −15.27 pp (95% interval −16.24 to −14.29 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −18.54 pp (95% interval −27.30 to −10.80 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P3 · Label template head against its own read-off — TPL / read-off | asleep · implicit name | +4.720 (95% interval +4.410 to +5.113) (R 112.17 (82.27–166.15)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | asleep · explicit name | +2.566 (95% interval +2.179 to +3.010) (R 13.02 (8.84–20.30)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | N1, N2 or N3 sleep · explicit name | +1.080 (95% interval +0.890 to +1.267) (R 2.95 (2.44–3.55)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · implicit name | +2.409 (95% interval +2.208 to +2.609) (R 11.12 (9.10–13.58)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · explicit name | +0.888 (95% interval +0.769 to +1.024) (R 2.43 (2.16–2.78)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | asleep · new descriptions — 10 held-out wordings, resampled | +4.707 (95% interval +4.279 to +5.154) (R 110.72 (72.15–173.16)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | N1, N2 or N3 sleep · new descriptions — 10 held-out wordings, resampled | +2.223 (95% interval +1.622 to +2.649) (R 9.24 (5.06–14.15)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | light sleep · new descriptions — 10 held-out wordings, resampled | +1.707 (95% interval +1.240 to +2.056) (R 5.51 (3.46–7.81)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, implicit names | −0.391 (95% interval −0.470 to −0.311) | SHUF higher | Difference only: no margin | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, explicit names | +0.667 (95% interval +0.634 to +0.698) | TPL higher | Difference only: no margin | Headroom gate passed | ### S2 · OpenBMI motor imagery, frozen CBraMod features · 51 people, three seeds Motor imagery has only two three-cycles, so two of the three shuffled-template seeds share one derangement. The unseen “imagining a hand movement” came closest to its read-off: a difference, but within the ratio margin. | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 3-way | −0.49 pp (95% interval −1.07 to +0.02 pp) | No difference shown | Equivalent within ±2 pp | No canary declared for this task: not run | | P1 · Description − question number — DESC − ID | seen, 3-way | −0.54 pp (95% interval −1.08 to +0.05 pp) | No difference shown | Equivalent within ±2 pp | No canary declared for this task: not run | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | +0.05 pp (95% interval −0.10 to +0.20 pp) | No difference shown | Equivalent within ±2 pp | No canary declared for this task: not run | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | +0.22 pp (95% interval −0.11 to +0.55 pp) | No difference shown | Equivalent within ±2 pp | No canary declared for this task: not run | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −3.49 pp (95% interval −6.63 to −1.42 pp) | DESC(train) higher | 2 pp margin not met | No canary declared for this task: not run | | P3 · Label template head against its own read-off — TPL / read-off | imagining a hand movement · implicit name | +0.076 (95% interval +0.050 to +0.107) (R 1.08 (1.05–1.11)) | Read-off leaves less error | Equivalent within the 1.25 ratio margin | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | imagining a hand movement · new descriptions — 10 held-out wordings, resampled | +0.088 (95% interval +0.048 to +0.138) (R 1.09 (1.05–1.15)) | Read-off leaves less error | Equivalent within the 1.25 ratio margin | Headroom gate passed | ### S3 · REVE-L features, BETA · 70 people, one seed · declared likely inconclusive REVE-L features keep the stimulus phase, so their neighbour 3-way tests frequency and phase together. | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 32-way | −2.76 pp (95% interval −3.51 to −2.07 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 32-way | −2.56 pp (95% interval −3.24 to −1.90 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −4.48 pp (95% interval −7.66 to −2.16 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −2.16 pp (95% interval −2.44 to −1.88 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −5.05 pp (95% interval −6.50 to −3.77 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, 8-way | −65.32 pp (95% interval −69.18 to −61.17 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − average of the two neighbouring seen frequencies — TPL − NN(TPL) | unseen frequency, 8-way | +5.81 pp (95% interval +4.54 to +7.04 pp) | TPL higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, 8-way | +1.92 pp (95% interval +0.85 to +2.92 pp) | TPL higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, neighbour 3-way | −38.75 pp (95% interval −41.62 to −35.60 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, neighbour 3-way | +35.22 pp (95% interval +32.91 to +37.42 pp) | TPL higher | Difference only: no margin | Headroom gate passed | ### S4 · text encoder all-MiniLM-L6-v2, BETA, a plain spectrum · 70 people, one seed · declared likely inconclusive | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 32-way | −6.73 pp (95% interval −7.75 to −5.68 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 32-way | −11.58 pp (95% interval −13.16 to −9.97 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −3.95 pp (95% interval −5.11 to −2.86 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −4.90 pp (95% interval −5.53 to −4.25 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −5.49 pp (95% interval −6.48 to −4.60 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, 8-way | −50.10 pp (95% interval −53.06 to −46.73 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − average of the two neighbouring seen frequencies — TPL − NN(TPL) | unseen frequency, 8-way | −7.68 pp (95% interval −8.82 to −6.50 pp) | NN(TPL) higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, 8-way | −10.84 pp (95% interval −12.53 to −9.08 pp) | NUM higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, neighbour 3-way | −61.85 pp (95% interval −65.13 to −58.31 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, neighbour 3-way | −0.18 pp (95% interval −2.02 to +1.73 pp) | No difference shown | Difference only: no margin | Headroom gate passed | ### S4 · text encoder all-MiniLM-L6-v2, BETA, frozen CBraMod features · 70 people, one seed · declared likely inconclusive | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 32-way | −4.73 pp (95% interval −5.43 to −4.07 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 32-way | −8.84 pp (95% interval −10.12 to −7.60 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −2.89 pp (95% interval −3.73 to −2.08 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −3.09 pp (95% interval −3.57 to −2.63 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −2.98 pp (95% interval −3.67 to −2.34 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, 8-way | −61.55 pp (95% interval −65.21 to −57.49 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − average of the two neighbouring seen frequencies — TPL − NN(TPL) | unseen frequency, 8-way | −1.09 pp (95% interval −1.69 to −0.48 pp) | NN(TPL) higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, 8-way | −2.03 pp (95% interval −2.86 to −1.19 pp) | NUM higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, neighbour 3-way | −54.63 pp (95% interval −57.51 to −51.29 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, neighbour 3-way | +15.57 pp (95% interval +14.27 to +16.75 pp) | TPL higher | Difference only: no margin | Headroom gate passed | ### S5 · Chinese wordings with multilingual-e5-small, BETA, a plain spectrum · 70 people, one seed · declared likely inconclusive For this encoder each held-out frequency is a new token, not only a new combination of seen tokens (decision S0b-6). | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 32-way | −3.15 pp (95% interval −3.80 to −2.48 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 32-way | −4.94 pp (95% interval −5.69 to −4.18 pp) | ID higher | 2 pp margin not met | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −4.67 pp (95% interval −6.57 to −3.24 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −8.70 pp (95% interval −9.76 to −7.60 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −7.35 pp (95% interval −8.58 to −6.17 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, 8-way | −54.76 pp (95% interval −57.77 to −51.25 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − average of the two neighbouring seen frequencies — TPL − NN(TPL) | unseen frequency, 8-way | −15.81 pp (95% interval −17.68 to −13.94 pp) | NN(TPL) higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, 8-way | −15.51 pp (95% interval −17.58 to −13.44 pp) | NUM higher | Difference only: no margin | Headroom gate passed | | P4 · Label template − CCA, no training — TPL − CCA | unseen frequency, neighbour 3-way | −59.98 pp (95% interval −63.74 to −55.86 pp) | CCA higher | 2 pp margin not met | Headroom gate passed | | P4 · Label template − numeric code — TPL − NUM | unseen frequency, neighbour 3-way | +1.69 pp (95% interval −0.60 to +4.18 pp) | No difference shown | Difference only: no margin | Headroom gate passed | ### S8 · extrapolation to a contiguous held-out band at the top of the grid, BETA L0 · 70 people, one seed · declared likely inconclusive | Comparison | Difference, pp | Difference shown? | Margin, ±2 pp | | --- | --- | --- | --- | | TPL - CCA (8-way) | −65.37 pp (95% interval −69.21 to −61.40 pp) | CCA higher | 2 pp margin not met | | TPL - NUM (8-way) | +3.51 pp (95% interval +1.64 to +5.49 pp) | TPL higher | TPL non-inferior at 2 pp | | TPL - NN(TPL) (8-way) | Not defined (EA-1) | | | | TPL - CCA (3-way) | −50.19 pp (95% interval −54.05 to −46.25 pp) | CCA higher | 2 pp margin not met | ### S8 · extrapolation to a contiguous held-out band at the top of the grid, BETA L1 · 70 people, one seed · declared likely inconclusive | Comparison | Difference, pp | Difference shown? | Margin, ±2 pp | | --- | --- | --- | --- | | TPL - CCA (8-way) | −67.47 pp (95% interval −71.22 to −63.56 pp) | CCA higher | 2 pp margin not met | | TPL - NUM (8-way) | +0.88 pp (95% interval −0.70 to +2.47 pp) | No difference shown | TPL non-inferior at 2 pp | | TPL - NN(TPL) (8-way) | Not defined (EA-1) | | | | TPL - CCA (3-way) | −48.41 pp (95% interval −52.10 to −44.40 pp) | CCA higher | 2 pp margin not met | In the band the neighbour 3-way competitors are themselves held-out questions (EA-1). For bge-small-en-v1.5, many of these texts carry a number token never seen in training. ### S12 · Wearable-102, heads trained on all of BETA’s frequencies and asked about off-grid ones · 102 people, three seeds The off-grid frequencies have two decimals, whose hundredths never occur in BETA’s training texts, so the text arms face new tokens; the numeric code is the clean comparison. Nine fits, three arms by three seeds, fixed before the freeze. | Comparison | Difference, pp | Difference shown? | Margin, ±2 pp | | --- | --- | --- | --- | | Dry electrodes | | | | | TPL - CCA (12-way) | −48.18 pp (95% interval −52.36 to −43.90 pp) | CCA higher | 2 pp margin not met | | NUM - CCA (12-way) | −42.16 pp (95% interval −46.05 to −38.38 pp) | CCA higher | 2 pp margin not met | | TPL - NUM (12-way) | −6.02 pp (95% interval −7.32 to −4.73 pp) | NUM higher | 2 pp margin not met | | DESC - CCA (12-way) | −48.52 pp (95% interval −52.99 to −44.02 pp) | CCA higher | 2 pp margin not met | | Wet electrodes | | | | | TPL - CCA (12-way) | −62.92 pp (95% interval −66.16 to −59.66 pp) | CCA higher | 2 pp margin not met | | NUM - CCA (12-way) | −56.11 pp (95% interval −59.16 to −52.83 pp) | CCA higher | 2 pp margin not met | | TPL - NUM (12-way) | −6.82 pp (95% interval −8.31 to −5.33 pp) | NUM higher | 2 pp margin not met | | DESC - CCA (12-way) | −63.56 pp (95% interval −66.68 to −60.31 pp) | CCA higher | 2 pp margin not met | ### S10, S11 and S13 · BETA L0: other arms against CCA and FBCCA, and held-out wordings · 70 people, three seeds | Comparison | Difference, pp | Difference shown? | Margin, ±2 pp | | --- | --- | --- | --- | | S10 DESC - CCA (8-way) | −52.29 pp (95% interval −55.41 to −49.00 pp) | CCA higher | 2 pp margin not met | | S10 DESC - CCA (3-way) | −62.72 pp (95% interval −66.34 to −58.98 pp) | CCA higher | 2 pp margin not met | | S10 DESC - NN(DESC) (8-way) | −10.80 pp (95% interval −12.15 to −9.39 pp) | NN(DESC) higher | 2 pp margin not met | | S10 NUM - CCA (8-way) | −39.44 pp (95% interval −42.25 to −36.48 pp) | CCA higher | 2 pp margin not met | | S10 TPL - NN(ID) (8-way) | −15.63 pp (95% interval −17.48 to −13.62 pp) | NN(ID) higher | 2 pp margin not met | | S10 TPL - SHUF (8-way) | +19.50 pp (95% interval +17.45 to +21.62 pp) | TPL higher | TPL non-inferior at 2 pp | | S11 TPL - FBCCA (8-way) | −61.68 pp (95% interval −63.72 to −59.69 pp) | FBCCA higher | 2 pp margin not met | | S11 TPL - FBCCA (3-way) | −73.35 pp (95% interval −75.61 to −70.93 pp) | FBCCA higher | 2 pp margin not met | | S13 TPL held-out frames - CCA (8-way) | −53.71 pp (95% interval −56.73 to −50.27 pp) | CCA higher | 2 pp margin not met | | S13 DESC held-out descriptions - CCA (8-way) | −54.01 pp (95% interval −57.07 to −50.83 pp) | CCA higher | 2 pp margin not met | S9 and S11 · BETA L0: accuracy on the seen and unseen frequencies in one test, and FBCCA | Arm | Seen frequencies | Unseen frequencies | | --- | --- | --- | | S9 TPL | 23.4% (95% interval 21.0%–25.7%) | 4.4% (95% interval 4.0%–4.8%) | | S9 DESC | 24.0% (95% interval 21.5%–26.5%) | 4.2% (95% interval 3.8%–4.7%) | | S9 SHUF | 23.4% (95% interval 20.9%–25.7%) | 1.1% (95% interval 1.0%–1.4%) | | S9 NUM | 20.7% (95% interval 18.5%–22.9%) | 11.5% (95% interval 10.2%–12.7%) | | One-output head, neighbour read-off (nn8) | 44.3% (95% interval 40.9%–47.5%) | | | FBCCA, seen 32-way | 81.1% (95% interval 77.0%–84.8%) | | | FBCCA, unseen 8-way | 90.4% (95% interval 87.7%–92.8%) | | | FBCCA, neighbour 3-way | 92.7% (95% interval 90.9%–94.4%) | | ### S10, S11 and S13 · BETA L1: other arms against CCA and FBCCA, and held-out wordings · 70 people, three seeds | Comparison | Difference, pp | Difference shown? | Margin, ±2 pp | | --- | --- | --- | --- | | S10 DESC - CCA (8-way) | −62.26 pp (95% interval −66.01 to −58.41 pp) | CCA higher | 2 pp margin not met | | S10 DESC - CCA (3-way) | −54.93 pp (95% interval −58.01 to −51.72 pp) | CCA higher | 2 pp margin not met | | S10 DESC - NN(DESC) (8-way) | −1.27 pp (95% interval −2.08 to −0.46 pp) | NN(DESC) higher | 2 pp margin not met | | S10 NUM - CCA (8-way) | −59.48 pp (95% interval −62.92 to −55.87 pp) | CCA higher | 2 pp margin not met | | S10 TPL - NN(ID) (8-way) | +0.68 pp (95% interval −0.25 to +1.61 pp) | No difference shown | Equivalent within ±2 pp | | S10 TPL - SHUF (8-way) | +7.97 pp (95% interval +6.88 to +9.09 pp) | TPL higher | TPL non-inferior at 2 pp | | S11 TPL - FBCCA (8-way) | −71.41 pp (95% interval −73.53 to −69.17 pp) | FBCCA higher | 2 pp margin not met | | S11 TPL - FBCCA (3-way) | −65.66 pp (95% interval −67.61 to −63.62 pp) | FBCCA higher | 2 pp margin not met | | S13 TPL held-out frames - CCA (8-way) | −62.31 pp (95% interval −66.06 to −58.27 pp) | CCA higher | 2 pp margin not met | | S13 DESC held-out descriptions - CCA (8-way) | −62.58 pp (95% interval −66.15 to −58.59 pp) | CCA higher | 2 pp margin not met | S9 and S11 · BETA L1: accuracy on the seen and unseen frequencies in one test, and FBCCA | Arm | Seen frequencies | Unseen frequencies | | --- | --- | --- | | S9 TPL | 18.8% (95% interval 16.7%–21.1%) | 2.5% (95% interval 2.2%–2.8%) | | S9 DESC | 18.9% (95% interval 16.9%–21.2%) | 2.5% (95% interval 2.3%–2.8%) | | S9 SHUF | 18.3% (95% interval 16.4%–20.4%) | 1.6% (95% interval 1.3%–1.9%) | | S9 NUM | 11.5% (95% interval 10.4%–12.7%) | 4.3% (95% interval 3.8%–4.8%) | | One-output head, neighbour read-off (nn8) | 17.6% (95% interval 16.6%–18.8%) | | | FBCCA, seen 32-way | 81.1% (95% interval 77.0%–84.8%) | | | FBCCA, unseen 8-way | 90.4% (95% interval 87.7%–92.8%) | | | FBCCA, neighbour 3-way | 92.7% (95% interval 90.9%–94.4%) | | S9 asks about seen and unseen frequencies in one 40-way test; FBCCA, like CCA, needs no training. One seed for FBCCA. ### Secondary results on BOAS **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. ### S3 · REVE-L features, BOAS · 100 people, one seed · declared likely inconclusive | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 5-way | +0.27 pp (95% interval −0.36 to +0.86 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 5-way | +0.12 pp (95% interval −0.48 to +0.71 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −0.39 pp (95% interval −1.04 to +0.05 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −13.54 pp (95% interval −14.57 to −12.49 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −21.86 pp (95% interval −30.62 to −14.12 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P3 · Label template head against its own read-off — TPL / read-off | asleep · implicit name | +4.149 (95% interval +3.753 to +4.614) (R 63.39 (42.66–100.85)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | asleep · explicit name | +2.647 (95% interval +2.274 to +3.070) (R 14.11 (9.72–21.55)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | N1, N2 or N3 sleep · explicit name | +1.317 (95% interval +1.163 to +1.483) (R 3.73 (3.20–4.41)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · implicit name | +2.730 (95% interval +2.540 to +2.929) (R 15.33 (12.67–18.71)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · explicit name | +1.872 (95% interval +1.691 to +2.057) (R 6.50 (5.43–7.82)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | asleep · new descriptions — 10 held-out wordings, resampled | +4.092 (95% interval +3.565 to +4.610) (R 59.84 (35.35–100.44)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | N1, N2 or N3 sleep · new descriptions — 10 held-out wordings, resampled | +1.829 (95% interval +1.250 to +2.292) (R 6.23 (3.49–9.90)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | light sleep · new descriptions — 10 held-out wordings, resampled | +1.774 (95% interval +1.162 to +2.236) (R 5.89 (3.20–9.35)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, implicit names | −0.585 (95% interval −0.612 to −0.555) | SHUF higher | Difference only: no margin | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, explicit names | +0.626 (95% interval +0.602 to +0.648) | TPL higher | Difference only: no margin | Headroom gate passed | ### S4 · text encoder all-MiniLM-L6-v2, BOAS, frozen CBraMod features · 100 people, one seed · declared likely inconclusive | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 5-way | +0.19 pp (95% interval −0.37 to +0.75 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 5-way | +0.01 pp (95% interval −0.64 to +0.72 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −1.72 pp (95% interval −4.31 to −0.09 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −14.04 pp (95% interval −15.29 to −12.68 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −25.74 pp (95% interval −34.29 to −16.26 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P3 · Label template head against its own read-off — TPL / read-off | asleep · implicit name | +3.798 (95% interval +3.385 to +4.273) (R 44.60 (29.53–71.76)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | asleep · explicit name | +1.325 (95% interval +0.999 to +1.715) (R 3.76 (2.71–5.56)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | N1, N2 or N3 sleep · explicit name | +0.882 (95% interval +0.717 to +1.047) (R 2.41 (2.05–2.85)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · implicit name | +2.338 (95% interval +2.134 to +2.552) (R 10.36 (8.45–12.83)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · explicit name | +1.796 (95% interval +1.598 to +2.000) (R 6.03 (4.94–7.39)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | asleep · new descriptions — 10 held-out wordings, resampled | +3.069 (95% interval +2.508 to +3.601) (R 21.52 (12.28–36.65)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | N1, N2 or N3 sleep · new descriptions — 10 held-out wordings, resampled | +1.118 (95% interval +0.793 to +1.449) (R 3.06 (2.21–4.26)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | light sleep · new descriptions — 10 held-out wordings, resampled | +1.070 (95% interval +0.840 to +1.333) (R 2.92 (2.32–3.79)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, implicit names | −0.626 (95% interval −0.653 to −0.596) | SHUF higher | Difference only: no margin | Headroom gate passed | | P5 · Label template − shuffled templates — TPL − SHUF | unseen questions, explicit names | +0.572 (95% interval +0.550 to +0.592) | TPL higher | Difference only: no margin | Headroom gate passed | ### S5 · Chinese wordings with multilingual-e5-small, BOAS, frozen CBraMod features · 100 people, one seed · declared likely inconclusive | Comparison | Question | Difference: pp, log R or AUROC | Difference shown? | Margin | Gate | | --- | --- | --- | --- | --- | --- | | P1 · Label template − question number — TPL − ID | seen, 5-way | +0.34 pp (95% interval −0.29 to +0.99 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P1 · Description − question number — DESC − ID | seen, 5-way | −0.28 pp (95% interval −0.93 to +0.42 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · New sentence frame − training wording — TPL(a1) − TPL(train) | seen, reworded — 12 held-out wordings, resampled | −0.70 pp (95% interval −1.83 to +0.08 pp) | No difference shown | Equivalent within ±2 pp | Passes the triviality canary | | P2 · Synonym − training wording — TPL(a2) − TPL(train) | seen, reworded — for these three fixed synonyms | −12.51 pp (95% interval −13.73 to −11.27 pp) | TPL(train) higher | 2 pp margin not met | Passes the triviality canary | | P2 · New description − training descriptions — DESC(a3) − DESC(train) | seen, reworded — 10 held-out wordings, resampled | −25.81 pp (95% interval −33.97 to −17.73 pp) | DESC(train) higher | 2 pp margin not met | Passes the triviality canary | | P3 · Label template head against its own read-off — TPL / read-off | asleep · implicit name | +3.826 (95% interval +3.424 to +4.277) (R 45.87 (30.69–72.04)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | asleep · explicit name | +1.558 (95% interval +1.216 to +1.950) (R 4.75 (3.37–7.03)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | N1, N2 or N3 sleep · explicit name | +0.737 (95% interval +0.584 to +0.898) (R 2.09 (1.79–2.45)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · implicit name | +2.322 (95% interval +2.122 to +2.532) (R 10.19 (8.34–12.57)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Label template head against its own read-off — TPL / read-off | light sleep · explicit name | +1.615 (95% interval +1.424 to +1.809) (R 5.03 (4.15–6.11)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | asleep · new descriptions — 10 held-out wordings, resampled | +2.715 (95% interval +2.127 to +3.244) (R 15.11 (8.39–25.63)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | N1, N2 or N3 sleep · new descriptions — 10 held-out wordings, resampled | +0.798 (95% interval +0.599 to +1.039) (R 2.22 (1.82–2.83)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | | P3 · Description head against its own read-off — DESC / read-off | light sleep · new descriptions — 10 held-out wordings, resampled | +0.943 (95% interval +0.773 to +1.123) (R 2.57 (2.17–3.07)) | Read-off leaves less error | 1.25 ratio margin not met | Headroom gate passed | ### S13 · held-out sentence frames against the read-off, BOAS · 100 people, three seeds | Comparison | log R, with R | Difference shown? | Margin, R < 1.25 | | --- | --- | --- | --- | | TPL asleep implicit (held-out frames) | +3.565 (95% interval +3.162 to +4.006) | Read-off leaves less error | 1.25 ratio margin not met | | TPL asleep explicit (held-out frames) | +1.976 (95% interval +1.594 to +2.392) | Read-off leaves less error | 1.25 ratio margin not met | | TPL NREM explicit (held-out frames) | +0.920 (95% interval +0.743 to +1.100) | Read-off leaves less error | 1.25 ratio margin not met | | TPL light implicit (held-out frames) | +2.233 (95% interval +2.022 to +2.443) | Read-off leaves less error | 1.25 ratio margin not met | | TPL light explicit (held-out frames) | +1.484 (95% interval +1.240 to +1.719) | Read-off leaves less error | 1.25 ratio margin not met | ### S6 · negation, BOAS: a boundary probe AUROC of each negated wording against the truth of the negated question, beside the AUROC of 1 − p(X), the head’s own answer to X turned around; three seeds. “Non-REM sleep” is scored against the truth of “not REM”. | Wording | Against the truth of the negated question | AUROC of 1 − p(X) | People | | --- | --- | --- | --- | | not W | 0.022 (95% interval 0.014–0.030) | 0.980 (95% interval 0.972–0.987) | 100 | | not N1 | 0.135 (95% interval 0.120–0.152) | 0.872 (95% interval 0.855–0.887) | 100 | | not N2 | 0.124 (95% interval 0.110–0.139) | 0.924 (95% interval 0.909–0.937) | 100 | | not N3 | 0.021 (95% interval 0.017–0.026) | 0.979 (95% interval 0.974–0.983) | 71 | | not REM | 0.049 (95% interval 0.038–0.061) | 0.956 (95% interval 0.944–0.967) | 100 | | non-REM sleep (implicit) | 0.048 (95% interval 0.037–0.060) | 0.956 (95% interval 0.944–0.967) | 100 | Every negated wording was answered as if it asked for X itself: against the truth of the negated question its AUROC lay far below 0.5, while turning the head’s answer to X around would have answered it well. Published with the results ## The wordings, the held-out frequencies and the shuffles All 787 texts the questions were asked with, in English and Chinese, for sleep, SSVEP and motor imagery: training and held-out sentence frames, label names and synonyms, descriptions, the names and descriptions of the unseen questions, and the negation frames. Two agents wrote them from a fixed brief before any score, the second writing every held-out description without seeing the first’s. The file also holds the leakage rules they passed, the held-out frequencies of every rotation and the extrapolation band, the declared derangements of the shuffled control, the numeric code of a frequency and the pinned text encoders. Text only, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). [The wordings, as JSON ↓](https://bci.report/data/questions-in-language-wordings.json) The sleep stages’ names in the training frames, and the three synonyms each was reworded with: | Stage | Training name | Synonyms (a2) | | --- | --- | --- | | W | wake | wakefulness, awake, stage W | | N1 | N1 sleep | stage 1 sleep, sleep stage one, stage N1 | | N2 | N2 sleep | stage 2 sleep, sleep stage two, stage N2 | | N3 | N3 sleep | slow-wave sleep, deep sleep, stage N3 | | REM | REM sleep | stage R sleep, paradoxical sleep, dream sleep | Methods & limits ## Limits - 'Unseen' means unseen by the EEG head; the text encoder has read these words, and every token of a held-out frequency also occurs in the training texts. - Held-out sleep questions are combinations of seen stages ('asleep' is the complement of one), reported as compositional beside the read-off. - Held-out SSVEP questions name a frequency, which CCA answers without training; interpolation tested, extrapolation secondary. - In the neighbour 3-way the unseen frequency competes against trained detectors for its neighbours; at L1 that test also involves stimulus phase. - Rewording and unseen results are separate rows and never pooled. - Wordings were written by agents from a fixed brief before scoring, not by users and not tuned. - Negation is a boundary probe; location and waveform-shape questions are not tested (no signal-level ground truth held). - Frozen features and a small head; one recipe; not a ranking of text or EEG encoders. - 35 pre-declared comparisons, not corrected for multiplicity. - Questions about negation, location or waveform shape need signal-level ground truth. Writing more prompts does not supply it. - Negation was not understood: the negated wordings ("not X", "non-REM sleep") scored as if they asked for X (AUROC 0.02-0.13 against the truth of the negated question; boundary probe S6, never a route sentence). - Implicit names of the unseen sleep unions ("asleep", "light sleep") ranked windows in reverse (TPL AUROC 0.317 and 0.244, below 0.5); only the explicit names ("N1, N2, N3 or REM sleep") carried the composition. - The neighbour 3-way at L1 involves stimulus phase as well as frequency; at L0 (a phase-free spectrum) it is the clean frequency test. - Unseen SSVEP questions compete against trained detectors for their neighbours (seen bias, stated). - The pre-run permutation canary failed twice (revisions 3 and 4) and was revised by the owner, who overrode revision 4's "no further revision"; revision 5 passed. Both failed records are kept and disclosed. - For multilingual-e5-small each held-out frequency is a new token (S0b-6); S8 and S12 also involve new tokens. - Secondary single-seed entries carry no training-run variance; EESM19 has 20 people; every EESM19 and single-seed entry was declared likely inconclusive. - BETA numbers use the published 2-s windows and folds; Wearable-102 numbers are computed from the 2026-09-20 Tsinghua author mirror and claim no physical unit. - BOAS: the three gaps and the attribution above; participants are pseudonymised in the public release; no result evaluates Bitbrain's headband or its automatic scoring. - The head here is not a language model, and none of these results shows that EEG understands language. ### Not run, and not reported - **revision-4 per-person centring diagnostic**: not run (revision 5 part A.2; no fit exists) - **Ying scorer agreement; factorial cognitive sets**: deferred (D10) - **E1 EEGNet trained jointly with the question vector**: deferred (D6) - **third text encoder; LEAF-style Q-Former heads; route-1 calibration on language heads; user-written wordings**: deferred (protocol.deferred) - **S8 TPL - NN(TPL)**: not defined (EA-1): the held-out band is contiguous, so no held-out frequency has two seen neighbours - **location and waveform-shape questions**: not tested: no signal-level ground truth held - **per-union paired contrasts beyond the P3 entries and the per-union levels**: not computed (W1) - **counts of people above or below zero on paired contrasts**: allowed for BOAS but not computed by stage 2; not added - **per-person distributions, per-fold values, per-window scores**: private by design ### Not published - Per-person, per-night, per-recording and per-fold values, per-window scores, features, head weights and fold assignments. - The values of the pre-run checks, which are published as pass or fail and as counts. - Measured compute, timed on a shared GPU. - Demographics, recording dates and clock times, and any image of a real recording. ### The text encoders - [BAAI/bge-small-en-v1.5 ↗](https://huggingface.co/BAAI/bge-small-en-v1.5): primary · MIT · `5c38ec7` - [sentence-transformers/all-MiniLM-L6-v2 ↗](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2): secondary, S4 · Apache-2.0 · `1110a24` - [intfloat/multilingual-e5-small ↗](https://huggingface.co/intfloat/multilingual-e5-small): Chinese wordings, S5 · MIT · `614241f` Frozen at pinned commits and never redistributed. ### The recordings ### BETA Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. [Source ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ### BOAS, the Bitbrain Open Access Sleep dataset Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. Published with three stated gaps, printed beside its results above. [Source ↗](https://openneuro.org/datasets/ds005555/versions/1.1.3) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) ### EESM19 Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. [Source ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) · [CC0-1.0 declared by upstream and mirror](https://creativecommons.org/publicdomain/zero/1.0/) ### OpenBMI motor imagery Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. [Source ↗](https://doi.org/10.5524/100542) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) ### Wearable SSVEP BCI dataset (dry and wet electrodes) Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces. Sensors, 21(4), 1256. https://doi.org/10.3390/s21041256 Computed from the official Tsinghua BCI Lab author mirror snapshot acquired 2026-09-20; the mirror itself has no separate DOI or version number. [Source ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) Features: frozen [CBraMod](https://bci.report/methods/cbramod/) (primary) and frozen [REVE](https://bci.report/methods/reve/) Large @ 317531c7, REVE-L (secondary). Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Large @ 317531c7). No model’s authors endorse these results. [REVE, arXiv:2510.21585 ↗](https://arxiv.org/abs/2510.21585) ### Methods and sources these results draw on - [Yu & Yao, 2026 · Visual Jev: Accurate and Efficient Decisions from Shared Visual Context ↗](https://arxiv.org/abs/2609.25845) - [Perez et al., 2017 · FiLM: Visual Reasoning with a General Conditioning Layer ↗](https://arxiv.org/abs/1709.07871) - [Lin, Zhang, Wu & Gao, 2006 · Frequency Recognition Based on CCA for SSVEP-Based BCIs, IEEE TBME 53:2610–2614 ↗](https://doi.org/10.1109/TBME.2006.886577) - [Chen, Wang, Gao, Jung & Gao, 2015 · Filter bank CCA for implementing a high-speed SSVEP-based BCI, J. Neural Eng. 12:046008 ↗](https://doi.org/10.1088/1741-2560/12/4/046008) - [Wallace et al., 2019 · Do NLP Models Know Numbers? Probing Numeracy in Embeddings, EMNLP 2019 ↗](https://aclanthology.org/D19-1534/) - [Weller, Lawrie & Van Durme, 2024 · NevIR: Negation in Neural Information Retrieval, EACL 2024 ↗](https://aclanthology.org/2024.eacl-long.139/) - [Xian, Lampert, Schiele & Akata · Zero-Shot Learning — A Comprehensive Evaluation of the Good, the Bad and the Ugly, TPAMI ↗](https://arxiv.org/abs/1707.00600) Data source: [questions-in-language-update.json](https://bci.report/data/questions-in-language-update.json) · schema bci-report-questions-in-language-update-v1. Wordings: [questions-in-language-wordings.json](https://bci.report/data/questions-in-language-wordings.json) · schema bci-report-questions-in-language-wordings-v1. A longer write-up of the BETA results is in Towards AI: [Can an EEG decoder answer a question it was never trained on? ↗](https://pub.towardsai.net/can-an-eeg-decoder-answer-a-question-it-was-never-trained-on-a9b0b7525b39) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Can an EEG model answer questions asked in words?* https://bci.report/topics/questions-in-language/ Figures from release [`questions-in-language-update-20261008`](https://bci.report/releases/#questions-in-language-update-20261008) (2026-10-08). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/questions-in-language/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # What lets a decoder answer an SSVEP frequency it never trained on? Route 3 follow-ups · SSVEP frequencies the decoder never trained on Fourteen follow-ups to route 3 each test one idea: how the question is asked, the head, stimulus phase, the matching computation, learned corrections to CCA, cross-spectral and eigenvalue readouts, pooling CCA’s per-harmonic statistics, filter-bank CCA with shorter windows and fewer electrodes, and weights on the harmonics. Thirteen run on its BETA cohort and five on Wearable-102, one of them only there. - [Reviewed aggregate JSON ↓](https://bci.report/data/new-ssvep-frequencies-update.json) - Measured on: [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) ## Short answer On BETA’s held-out frequencies, temporal CCA with a small learned correction reached **86.15%** against **80.90%** for CCA alone, choosing among **8** frequencies (**+5.24 pp**, **+3.91** to **+6.62**; **70** people). Run unchanged on Wearable-102, the same models changed CCA’s score by **+0.07 pp** on dry electrodes (**−0.71** to **+0.91**) and **−0.61 pp** on wet (**−1.48** to **+0.23**). Learned heads that read reference spectra or stimulus phase stayed far below CCA. On the same held-out task, filter-bank CCA, which needs no training, reached **90.21%** against **80.90%** for CCA (**+9.31 pp**, **+6.84** to **+11.71**), and the untrained harmonic sum came out nearly level with the learned eigenvalue head (the head minus the sum: **+0.07 pp**, **−0.45** to **+0.62**). Tasks, cohort and names ## How the follow-ups are set up These follow-ups answer questions from [a reader’s comments on Hugging Face](https://huggingface.co/posts/Twu31/109502481720473). The tasks the cards report, with the chance level of each. | Task | Candidates | Chance | Cards | What it measures | | --- | --- | --- | --- | --- | | Held-out frequencies | 8 | 12.50% | 1 · 2 · 3 · 4 · 5 · 6 · 8 · 9 · 10 · 11 · 12 · 13 · 14 | Accuracy among the frequencies left out of training the decoder in that rotation. | | Seen frequencies, balanced | 32 | 3.13% | 1 · 2 · 3 · 4 · 5 · 6 · 8 · 9 · 10 · 11 · 13 · 14 | Balanced accuracy among the frequencies the decoder trained on. | | Core BETA protocols | 40 | 2.50% | [BETA, eight channels](https://bci.report/protocols/beta-8ch/) · [BETA, four channels](https://bci.report/protocols/beta-4ch/) | The core matrix’s task, reported on its own protocol pages. | | Wearable-102 | 12 | 8.33% | 7 · 10 · 12 · 13 · 14 | Accuracy among Wearable-102’s targets; dry and wet headbands are scored separately. | | Question number on held-out frequencies | — | not defined | — | A question number exists only for a frequency seen in training, so it cannot ask about a held-out one. | Cards 1 to 6: BETA, 70 people, 11,200 two-second windows, 7 participant folds × 5 frequency rotations (32 seen, 8 held out); trained heads use 3 seeds. These follow-ups reuse the cohort [route 3](https://bci.report/topics/questions-in-language/) already studied. “Unseen” means unseen by the trained part of the decoder. Cards 8 to 14: the same BETA cohort, 70 people, and on cards 10, 12, 13 and 14 Wearable-102’s 102 people, dry and wet. They are repeated analyses of data these pages already studied, not external replication. Each card names its windows, set-ups and interval method. The names this page uses, their export codes and the cards that use them The names this page uses, with the codes the data file uses for them. | Name on this page | Export code | Cards | | --- | --- | --- | | question number | `ID` | 1 · 2 · 4 | | numeric code | `NUM` | 1 · 2 · 4 | | label template | `TPL` | 1 · 2 | | reference spectrum | `REF` | 1 · 2 · 3 · 4 | | reference with stimulus phase | `REFc-0` | 2 · 3 | | fixed spectrum match (no training) | `TEMPLATE` | 1 | | plain spectrum | `L0` | 1 · 2 · 3 · 4 | | complex spectrum | `L0c` | 2 · 3 · 4 | | frozen CBraMod features | `L1` | 1 | | FiLM head | `FS` | 1 · 2 · 3 · 4 | | matching head | `MATCH` | 2 · 4 | | CCA (no training) | `CCA` · `CCA-RAW` · `CCA-RAW-H3` — CCA-RAW-H3 in card 11’s export: the same scores at 2 s and eight electrodes | 1 · 2 · 4 · 5 · 6 · 8 · 9 · 10 · 11 | | Hann-tapered CCA | `CCA-HANN-C` · `CCA-HANN-U` | 5 | | cropped-spectrum match (no training) | `SUBSPACE-CROP` · `CCA-CROP` — one set of scores: SUBSPACE-CROP in card 5’s export, CCA-CROP in card 6’s | 5 · 6 · 8 | | channel-energy average | `CHANNEL-JOINT` · `CHANNEL-HARMONIC` | 5 · 8 | | shared ridge scorer | `SHARED-RIDGE` | 5 · 8 · 9 | | energy head | `ENERGY-FS` | 6 | | adapted head | `ADAPT-FS` | 6 | | CCA with learned correction | `RAW-RESIDUAL` | 6 · 7 | | cropped match with learned correction | `CROP-RESIDUAL` | 6 | | cropped-spectrum check (Wearable-102) | `CCA-CROP` — the code in card 7’s Wearable-102 export | 7 | | harmonic sum (no training) | `RAW-HARM-SUM` · `RAW-HARM-SUM-H3` · `S` · `H3-FLAT` · `flat` — one set of scores: S in card 12’s export, H3-FLAT in card 13’s, flat in card 14’s | 9 · 10 · 11 · 12 · 13 · 14 | | joint CCA | `JOINT-TOP` · `J` — an alias of CCA (no training) and matched CCA — J in card 12’s export; the same scores as CCA (no training) on BETA and matched CCA on Wearable-102 | 9 · 12 | | common-direction sum | `B` | 12 | | eigenvalue head | `EIG-RAW` · `EIG-RAW6-RIDGE` | 9 · 10 · 11 | | joint eigenvalue head | `EIG-JOINT` | 9 · 10 | | spectral eigenvalue head | `EIG-SPEC` | 9 | | spectral harmonic sum (no training) | `SPEC-HARM-SUM` | 9 | | coherence head | `COH-ABS` | 9 | | signed coherence head | `COH-SIGNED` | 9 | | top-three ridge | `RAW-LEAD3-RIDGE` | 10 · 11 | | harmonic max (no training) | `RAW-HARM-MAX` | 10 | | all-component sum (no training) | `RAW-HARM-ALLSUM` | 10 | | joint all-component sum (no training) | `JOINT-ALLSUM` | 10 | | cross-spectral ridge | `XS-RIDGE` | 8 · 9 | | whitened cross-spectral ridge | `XS-RIDGE-W` | 8 · 9 | | cross-spectral ridge with quadrature | `XS-RIDGE-Q` | 8 · 9 | | cross-spectral MLP | `XS-MLP` | 8 | | filter-bank CCA (seven sub-bands) | `FBCCA-M3-N7-H5` · `FBCCA-DECAY` — FBCCA-DECAY in card 13’s export: the same scores as card 11’s at 2 s and eight electrodes | 11 · 13 | Card 1 · how the question is asked ## Asking with the reference spectrum - Question: Asked with the spectrum of CCA’s own reference signals, does the head do better on frequencies it never trained on? - What changed: How each frequency is asked: the reference spectrum in place of the numeric code or the question number, read by the FiLM head on the same features and splits. - Matched control: The same FiLM head asked with the numeric code on held-out frequencies and with the question number on seen ones; CCA and a fixed spectrum match, neither of them trained, score the same windows. ### Result Asked with the spectrum of CCA’s own reference signals (sine and cosine at the frequency and its second and third harmonics), the head reached 39.00% (36.07%–41.78%) on frequencies it never trained on. The numeric code reached 41.46% and CCA 80.90%. Paired with the numeric code, the reference spectrum differed by −2.46 pp (−3.47 to −1.52). On the seen 32-way task it reached 31.48% balanced accuracy, against 32.06% for the question number: the paired difference, −0.57 pp (−1.06 to −0.05), is equivalent within the ±2-point margin and is also a small difference below zero. On frozen CBraMod features the reference spectrum was again below the numeric code on unseen frequencies (−2.35 pp, −3.16 to −1.57). On seen frequencies it was within the margin, with the direction unresolved (+0.44 pp, −0.01 to +0.92). This question changes the harmonics, the shape and size of the query and the size of the head together, so the result describes the recipe as a whole. The reference spectrum against each comparator: four planned contrasts on each feature set, in percentage points. | Contrast | Task | Role | Difference, 95% interval | Reading | | --- | --- | --- | --- | --- | | Plain spectrum · primary | | | | | | Reference spectrum − numeric code | Held-out | Primary | −2.46 pp (−3.47 to −1.52) | Below zero; read for a difference only | | Reference spectrum − question number | Seen, balanced | Primary | −0.57 pp (−1.06 to −0.05) | Below zero; equivalent within ±2 pp | | Reference spectrum − CCA (no training) | Held-out | Primary | −41.90 pp (−44.98 to −38.80) | Below zero; non-inferiority to CCA (−2 pp) not met | | Reference spectrum: held-out − m8 | Held-out against m8 | Primary | −18.18 pp (−19.64 to −16.69) | Below zero; cost not within ±2 pp of zero | | Frozen CBraMod features · secondary | | | | | | Reference spectrum − numeric code | Held-out | Secondary | −2.35 pp (−3.16 to −1.57) | Below zero; read for a difference only | | Reference spectrum − question number | Seen, balanced | Secondary | +0.44 pp (−0.01 to +0.92) | Direction unresolved; equivalent within ±2 pp | | Reference spectrum − CCA (no training) | Held-out | Secondary | −61.83 pp (−65.45 to −58.07) | Below zero; non-inferiority to CCA (−2 pp) not met | | Reference spectrum: held-out − m8 | Held-out against m8 | Secondary | −27.50 pp (−30.08 to −24.93) | Below zero; cost not within ±2 pp of zero | 95% intervals from 2,000 hierarchical paired bootstrap draws over people and fold-shared seed indices (seed 20261007). Four primary contrasts on the plain spectrum and four secondary on frozen CBraMod features, with no multiplicity correction. A fixed spectrum match against the same reference, with no training, reached 28.84% on unseen frequencies. Its scores do not depend on the rotation, so its generalisation cost is zero by construction, as is CCA’s. Every level behind card 1: held-out, m8, cost, seen, and the two read-offs Levels with their 95% intervals. Rows marked route 3 are route 3’s own runs, reused unchanged. The question number exists only for seen frequencies, so its held-out cells are not defined. CCA reads the raw multichannel window, which keeps the timing and phase the plain spectrum drops, and is the comparator for both feature sets. | Arm | Held-out | m8 | Cost, held-out − m8 | Seen, balanced | nn8 | u3 | | --- | --- | --- | --- | --- | --- | --- | | Plain spectrum · FiLM head · primary | | | | | | | | Reference spectrum | 39.00% (36.07%–41.78%) | 57.18% (53.38%–60.95%) | −18.18 pp (−19.64 to −16.69) | 31.48% (28.28%–34.62%) | 43.41% (40.04%–46.58%) | 19.61% (18.53%–20.63%) | | Numeric code — from route 3 | 41.46% (38.15%–44.68%) | 53.76% (49.87%–57.49%) | −12.30 pp (−13.42 to −11.17) | 23.67% (21.22%–26.09%) | 40.96% (37.65%–44.13%) | 20.57% (19.33%–21.78%) | | Question number — from route 3 | not defined | 58.58% (54.75%–62.33%) | not defined | 32.06% (28.79%–35.19%) | 44.32% (40.84%–47.57%) | not defined | | Label template — from route 3 | 28.69% (26.95%–30.55%) | 53.13% (49.57%–56.66%) | −24.44 pp (−26.44 to −22.26) | 26.10% (23.49%–28.59%) | 39.67% (36.78%–42.47%) | 19.33% (18.23%–20.47%) | | Frozen CBraMod features · FiLM head · secondary | | | | | | | | Reference spectrum | 19.07% (18.06%–20.22%) | 46.57% (43.37%–49.75%) | −27.50 pp (−30.08 to −24.93) | 24.68% (22.02%–27.43%) | 19.40% (18.35%–20.52%) | 24.49% (23.34%–25.69%) | | Numeric code — from route 3 | 21.42% (20.16%–22.62%) | 37.58% (35.01%–40.17%) | −16.16 pp (−17.88 to −14.47) | 13.86% (12.52%–15.22%) | 21.46% (20.20%–22.67%) | 11.42% (10.74%–12.12%) | | Question number — from route 3 | not defined | 46.08% (43.04%–49.16%) | not defined | 24.24% (21.61%–26.92%) | 18.28% (17.30%–19.30%) | not defined | | Label template — from route 3 | 18.96% (18.02%–19.88%) | 43.23% (40.20%–46.29%) | −24.27 pp (−26.72 to −21.85) | 20.75% (18.49%–23.09%) | 20.12% (18.98%–21.33%) | 27.03% (25.96%–28.05%) | | No training | | | | | | | | CCA (no training) — from route 3 | 80.90% (76.59%–84.85%) | 80.90% (76.59%–84.85%) | 0.00 pp (by construction) | 65.46% (59.66%–71.08%) | 66.66% (61.33%–71.56%) | 81.70% (78.49%–84.57%) | | Fixed spectrum match (no training) | 28.84% (26.46%–31.49%) | 28.84% (26.46%–31.49%) | 0.00 pp (by construction) | 12.85% (11.18%–14.58%) | 27.35% (25.09%–29.93%) | 31.99% (29.84%–34.08%) | **Limit**. Not run in this round: the matching (bilinear) head and a fundamental-only reference; [card 2](https://bci.report/topics/new-ssvep-frequencies/#head-and-features) runs both. Card 2 · head, features and question ## The head, the spectrum or the question - Question: What holds the head back: the head itself, the spectrum it reads, or the way a frequency is asked? And does a plain softmax do better on seen frequencies? - What changed: One named part per contrast: the head (a matching head in place of the FiLM head), the features (the complex spectrum, which keeps phase, in place of the plain spectrum), the question (the reference with stimulus phase, or the reference without its second and third harmonics) and, in a separate run, the output (a softmax with no question, or route 3’s head trained with a softmax loss). - Matched control: The same arm without that change, paired on the same people, folds, rotations and seeds; card 1’s reference spectrum, numeric code and CCA sit beside them as levels. ### Result A plain 32-way softmax on the same spectrum reached 35.90% seen balanced accuracy (32.30%–39.32%). The head with one output per question, trained with the same softmax loss, reached 31.45%; trained one-vs-rest, as in route 3, it reached 32.16%. The paired difference, −0.71 pp (−1.01 to −0.41), is small, below zero and inside the ±2-point margin. On the plain spectrum, a matching head in place of the FiLM head changed seen balanced accuracy by −3.60 pp with the reference spectrum (−4.37 to −2.73) and by +2.54 pp with question numbers (+1.74 to +3.37); the protocol counts a head gain as meaningful when its lower bound is above +2 points, and neither lower bound is. With question numbers, the complex spectrum, which also keeps phase, changed seen balanced accuracy by +18.03 pp for the FiLM head (+14.13 to +21.77), to 50.08%, and by +21.53 pp for the matching head (+17.25 to +25.63), to 56.14%. It also changes size, scaling and the head’s size, so these figures describe the feature recipe as a whole. On held-out frequencies, a question carrying each target’s stimulus phase changed the complex-spectrum FiLM head’s accuracy by +16.30 pp against the magnitude question (+13.36 to +19.32), to 24.97%, and the matching head’s by +15.95 pp (+13.38 to +18.45), to 22.47%. Read side by side, both stay below the reference spectrum on the plain spectrum (39.00%), the numeric code (41.46%) and CCA (80.90%); the matching head’s paired difference from CCA is −58.43 pp (−62.10 to −54.83). The stimulus phases come from the BETA authors’ corrigendum, converted to the cosine convention with zero latency; equal amplitudes across harmonics, with harmonic h at h times the stimulus phase, are modelling assumptions. They are the phases of the stimulus as displayed, and the question carries no phase measured on the scalp. Dropping the second and third harmonics from the reference spectrum changed the plain-spectrum FiLM head’s held-out accuracy by −13.27 pp (−15.04 to −11.45), to 25.73%, and its seen balanced accuracy by −3.12 pp (−3.65 to −2.57). Each contrast changes one named part; differences in percentage points. | Contrast | Task | Role | Difference, 95% interval | Reading | | --- | --- | --- | --- | --- | | The head · plain spectrum | | | | | | Matching head − FiLM head, asked with the reference spectrum | Seen, balanced | Primary | −3.60 pp (−4.37 to −2.73) | Below zero; meaningful head gain (lower bound above +2 pp) not met | | Matching head − FiLM head, asked with question numbers | Seen, balanced | Primary | +2.54 pp (+1.74 to +3.37) | Above zero; meaningful head gain (lower bound above +2 pp) not met | | The features · complex spectrum − plain spectrum | | | | | | FiLM head, asked with the reference spectrum | Seen, balanced | Primary | +15.81 pp (+11.91 to +19.51) | Above zero | | FiLM head, asked with question numbers | Seen, balanced | Primary | +18.03 pp (+14.13 to +21.77) | Above zero | | Matching head, asked with the reference spectrum | Seen, balanced | Primary | +26.19 pp (+21.85 to +30.42) | Above zero | | Matching head, asked with question numbers | Seen, balanced | Primary | +21.53 pp (+17.25 to +25.63) | Above zero | | The question · complex spectrum | | | | | | FiLM head: reference with stimulus phase − reference spectrum | Held-out | Primary | +16.30 pp (+13.36 to +19.32) | Above zero | | Matching head: reference with stimulus phase − reference spectrum | Held-out | Primary | +15.95 pp (+13.38 to +18.45) | Above zero | | Matching head: reference with stimulus phase − numeric code | Held-out | Primary | +12.14 pp (+9.36 to +14.82) | Above zero | | Matching head asked with the reference with stimulus phase − CCA (no training) | Held-out | Primary | −58.43 pp (−62.10 to −54.83) | Below zero; non-inferiority to CCA (−2 pp) not met | | The question · plain spectrum | | | | | | FiLM head: fundamental-only reference − reference spectrum | Held-out | Primary | −13.27 pp (−15.04 to −11.45) | Below zero | | FiLM head: fundamental-only reference − reference spectrum | Seen, balanced | Primary | −3.12 pp (−3.65 to −2.57) | Below zero | | The output · softmax run, plain spectrum | | | | | | One output per question: softmax training − one-vs-rest training | Seen, balanced | Primary | −0.71 pp (−1.01 to −0.41) | Below zero; equivalent within ±2 pp | | One output per question, softmax training − plain 32-way softmax | Seen, balanced | Secondary | −4.45 pp (−5.12 to −3.79) | Below zero; not equivalent within ±2 pp | 95% intervals from 2,000 hierarchical paired bootstrap draws over people and fold-shared seed indices (seed 20261007). Twelve planned contrasts in the main run, all primary, and one primary and one secondary in the softmax run, with no multiplicity correction. Every level behind card 2: held-out, m8, cost, seen, and the two read-offs Levels with their 95% intervals. Rows marked card 1 or route 3 are those runs, reused unchanged. A question number or a softmax exists only for seen frequencies, so their held-out cells are not defined. On the complex spectrum u3 also involves stimulus phase, because neighbouring targets differ in phase. | Arm | Held-out | m8 | Cost, held-out − m8 | Seen, balanced | nn8 | u3 | | --- | --- | --- | --- | --- | --- | --- | | Plain spectrum · FiLM head | | | | | | | | Reference spectrum — from card 1 | 39.00% (36.07%–41.78%) | 57.18% (53.38%–60.95%) | −18.18 pp (−19.64 to −16.69) | 31.48% (28.28%–34.62%) | 43.41% (40.04%–46.58%) | 19.61% (18.53%–20.63%) | | Fundamental-only reference | 25.73% (24.20%–27.24%) | 54.47% (50.84%–58.00%) | −28.74 pp (−31.17 to −26.16) | 28.36% (25.49%–31.13%) | 41.67% (38.48%–44.70%) | 18.37% (17.26%–19.47%) | | Question number — from route 3 | not defined | 58.58% (54.75%–62.33%) | not defined | 32.06% (28.79%–35.19%) | 44.32% (40.84%–47.57%) | not defined | | Numeric code — from route 3 | 41.46% (38.15%–44.68%) | 53.76% (49.87%–57.49%) | −12.30 pp (−13.42 to −11.17) | 23.67% (21.22%–26.09%) | 40.96% (37.65%–44.13%) | 20.57% (19.33%–21.78%) | | Label template — from route 3 | 28.69% (26.95%–30.55%) | 53.13% (49.57%–56.66%) | −24.44 pp (−26.44 to −22.26) | 26.10% (23.49%–28.59%) | 39.67% (36.78%–42.47%) | 19.33% (18.23%–20.47%) | | Plain spectrum · matching head | | | | | | | | Reference spectrum | 33.40% (31.10%–35.56%) | 51.86% (48.52%–55.13%) | −18.46 pp (−20.22 to −16.71) | 27.88% (25.04%–30.64%) | 40.36% (37.51%–43.27%) | 25.20% (24.27%–26.07%) | | Question number | not defined | 58.75% (55.03%–62.37%) | not defined | 34.60% (31.15%–37.91%) | 46.10% (42.76%–49.44%) | not defined | | Numeric code | 41.22% (37.84%–44.45%) | 53.98% (50.22%–57.55%) | −12.76 pp (−13.81 to −11.72) | 26.08% (23.40%–28.82%) | 40.99% (37.77%–44.03%) | 27.99% (26.30%–29.52%) | | Complex spectrum · FiLM head | | | | | | | | Reference spectrum | 8.67% (7.85%–9.61%) | 65.56% (61.57%–69.44%) | −56.89 pp (−61.14 to −52.56) | 47.29% (42.73%–51.92%) | 8.56% (7.66%–9.53%) | 25.46% (24.22%–26.73%) | | Reference with stimulus phase | 24.97% (22.60%–27.44%) | 66.96% (63.01%–70.83%) | −41.98 pp (−45.04 to −38.93) | 49.24% (44.54%–53.97%) | 7.44% (6.63%–8.26%) | 69.66% (65.97%–73.21%) | | Question number | not defined | 67.95% (63.95%–71.82%) | not defined | 50.08% (45.34%–54.78%) | 7.40% (6.58%–8.35%) | not defined | | Numeric code | 13.13% (12.14%–14.17%) | 50.71% (47.16%–54.37%) | −37.57 pp (−41.13 to −33.99) | 24.52% (22.01%–27.12%) | 13.41% (12.44%–14.41%) | 7.16% (6.63%–7.79%) | | Complex spectrum · matching head | | | | | | | | Reference spectrum | 6.52% (5.76%–7.34%) | 70.16% (66.13%–74.16%) | −63.64 pp (−68.11 to −59.06) | 54.07% (49.11%–59.13%) | 6.69% (5.88%–7.54%) | 23.99% (23.12%–24.83%) | | Reference with stimulus phase | 22.47% (20.46%–24.45%) | 70.00% (65.87%–74.03%) | −47.53 pp (−50.87 to −44.28) | 54.16% (49.25%–59.24%) | 6.75% (5.93%–7.58%) | 60.73% (58.00%–63.39%) | | Question number | not defined | 71.50% (67.56%–75.33%) | not defined | 56.14% (51.31%–61.08%) | 6.78% (5.90%–7.69%) | not defined | | Numeric code | 10.33% (9.39%–11.36%) | 50.03% (46.53%–53.66%) | −39.70 pp (−43.47 to −36.12) | 25.22% (22.72%–27.86%) | 10.68% (9.68%–11.79%) | 10.76% (9.85%–11.71%) | | Plain spectrum · no question, softmax run | | | | | | | | Plain 32-way softmax | not defined | — | — | 35.90% (32.30%–39.32%) | — | — | | One output per question, softmax training | not defined | — | — | 31.45% (28.28%–34.49%) | — | — | | One output per question, one-vs-rest training — from route 3 | not defined | — | — | 32.16% (28.93%–35.29%) | — | — | | No training | | | | | | | | CCA (no training) — from route 3 | 80.90% (76.59%–84.85%) | 80.90% (76.59%–84.85%) | 0.00 pp (by construction) | 65.46% (59.66%–71.08%) | 66.66% (61.33%–71.56%) | 81.70% (78.49%–84.57%) | **Limit**. Each arm is one fixed recipe: the matching head is one bilinear design, and the softmax run changes the whole training objective or several parts of the architecture at once. These contrasts compare the recipes; how much frequency information the plain spectrum holds stays open. Both runs were designed with route 3’s results known. Card 3 · stimulus phase ## A random delay with the phase-carrying query During training only, each window’s complex spectrum got a random delay common to all channels, drawn either from 0–125 ms or from 0 to one period of the training label’s frequency. The head was the FiLM head on the complex spectrum, asked with the reference with stimulus phase, the phase-carrying query (526,465 parameters, trained from scratch); queries and test windows were unchanged. On held-out frequencies the 0–125 ms arm reached 12.51% (11.93%–13.13%) and the one-period arm 13.13% (12.61%–13.68%). The same head without the delay reached 24.97% (22.60%–27.44%). Primary change, 0–125 ms minus no delay: −12.46 pp (−14.84 to −10.25). Seen balanced accuracy fell from 49.24% to 15.70% and 15.28%. Card 3’s four arms: accuracy with 95% intervals. | Arm | Head · features · query | Held-out frequencies | Seen frequencies, balanced | | --- | --- | --- | --- | | 0–125 ms delay — `COMMON-125MS` | FiLM head · complex spectrum · reference with stimulus phase | 12.51% (11.93%–13.13%) | 15.70% (13.71%–17.88%) | | one-period delay — `COMMON-ONE-PERIOD` | FiLM head · complex spectrum · reference with stimulus phase | 13.13% (12.61%–13.68%) | 15.28% (13.83%–16.69%) | | no delay (matched control) — `L0c-FS-REFc-0` | FiLM head · complex spectrum · reference with stimulus phase | 24.97% (22.60%–27.44%) | 49.24% (44.54%–53.97%) | | magnitude-reference head on the plain spectrum — `L0-FS-REF` | FiLM head · plain spectrum · reference spectrum | 39.00% (36.07%–41.78%) | 31.48% (28.28%–34.62%) | Only the 0–125 ms minus no-delay contrast is primary; the others are descriptive. 95% intervals from 2,000 hierarchical paired bootstrap draws over people and training seeds (seed 20261007), conditional on the fixed person folds and three seeds; the five fixed rotations are averaged first. None of the 6 contrasts is adjusted for multiplicity. The delay multiplies already-windowed coefficients by one phase ramp, which approximates an onset shift. The one-period range is drawn from the training label’s frequency; the 0–125 ms range ignores the label and equals one period only at the lowest stimulus frequency, so its phase spread differs by target. BETA assigns each target its own stimulus phase, in equal steps between neighbouring frequencies, so the delay also removes information tied to the label. The magnitude-reference head on the plain spectrum (39.00%) differs in features, query and size. All 4 randomized-label checks passed (40 fits); a training-only shortcut through the label-dependent delay remains possible. The 6 paired contrasts Each contrast is a paired difference, left minus right, in percentage points. | Contrast | Task | Comparison | Role | Estimate | 95% interval | | --- | --- | --- | --- | --- | --- | | `P1` | Held-out | 0–125 ms delay minus no delay | primary | −12.46 pp | −14.84 to −10.25 | | `S1` | Held-out | one-period delay minus no delay | descriptive | −11.85 pp | −14.29 to −9.51 | | `S2` | Held-out | one-period delay minus magnitude-reference head on the plain spectrum | descriptive | −25.87 pp | −28.64 to −22.99 | | `S3` | Held-out | one-period delay minus 0–125 ms delay | descriptive | +0.62 pp | −0.13 to +1.40 | | `S4` | Seen, balanced | 0–125 ms delay minus no delay | descriptive | −33.54 pp | −37.84 to −29.45 | | `S5` | Seen, balanced | one-period delay minus no delay | descriptive | −33.97 pp | −38.12 to −29.88 | Card 4 · stimulus phase ## A random phase ramp with magnitude queries During training only, each window’s complex spectrum got a random delay of up to one stimulus period, applied as one phase ramp across all channels. Queries and test windows were unchanged. With the magnitude reference as the query, the FiLM head reached 30.56% on held-out frequencies (27.81%–33.22%). Against the same head and query on the plain spectrum the change was −8.45 pp (primary; −10.14 to −6.69); against the same head and query on the complex spectrum without the ramp, which reached 8.67%, it was +21.88 pp (secondary; +18.78 to +25.00). Seen balanced accuracy fell from 47.29% to 15.03%. The matching head ended at chance on both tasks. CCA reached 80.90% on the same held-out task. The bet in the thread was that held-out accuracy would clear 39%; the FiLM head reached 30.56% with the magnitude reference and 28.10% with the numeric code. Card 4’s arms: accuracy with 95% intervals. A dash marks a score that is not defined: a question number cannot ask about a held-out frequency. | Head · query | Held-out frequencies | m8, the matched comparison | Seen frequencies, balanced | | --- | --- | --- | --- | | Complex spectrum, trained with the ramp | | | | | FiLM head · reference spectrum — `FS-REF` | 30.56% (27.81%–33.22%) | 42.86% (39.29%–46.25%) | 15.03% (13.50%–16.51%) | | FiLM head · numeric code — `FS-NUM` | 28.10% (25.48%–30.58%) | 35.35% (32.14%–38.39%) | 10.92% (9.72%–12.11%) | | FiLM head · question number — `FS-ID` | — | 14.03% (13.62%–14.47%) | 3.70% (3.49%–3.90%) | | matching head · reference spectrum — `MATCH-REF` | 12.93% (12.31%–13.51%) | 12.33% (11.98%–12.65%) | 3.13% (2.95%–3.30%) | | matching head · numeric code — `MATCH-NUM` | 12.42% (11.87%–12.99%) | 12.37% (11.92%–12.83%) | 3.03% (2.81%–3.22%) | | matching head · question number — `MATCH-ID` | — | 12.01% (11.38%–12.65%) | 3.15% (2.89%–3.42%) | | Complex spectrum, no ramp (matched controls) | | | | | FiLM head · reference spectrum — `L0c-FS-REF` | 8.67% (7.85%–9.61%) | 65.56% (61.57%–69.44%) | 47.29% (42.73%–51.92%) | | FiLM head · numeric code — `L0c-FS-NUM` | 13.13% (12.14%–14.17%) | 50.71% (47.16%–54.37%) | 24.52% (22.01%–27.12%) | | FiLM head · question number — `L0c-FS-ID` | — | 67.95% (63.95%–71.82%) | 50.08% (45.34%–54.78%) | | matching head · reference spectrum — `L0c-MATCH-REF` | 6.52% (5.76%–7.34%) | 70.16% (66.13%–74.16%) | 54.07% (49.11%–59.13%) | | matching head · numeric code — `L0c-MATCH-NUM` | 10.33% (9.39%–11.36%) | 50.03% (46.53%–53.66%) | 25.22% (22.72%–27.86%) | | matching head · question number — `L0c-MATCH-ID` | — | 71.50% (67.56%–75.33%) | 56.14% (51.31%–61.08%) | | Plain spectrum, no ramp | | | | | FiLM head · reference spectrum — `L0-FS-REF` | 39.00% (36.07%–41.78%) | 57.18% (53.38%–60.95%) | 31.48% (28.28%–34.62%) | | FiLM head · numeric code — `L0-FS-NUM` | 41.46% (38.15%–44.68%) | 53.76% (49.87%–57.49%) | 23.67% (21.22%–26.09%) | | No training | | | | | CCA (no training) — `CCA` | 80.90% (76.59%–84.85%) | 80.90% (76.59%–84.85%) | 65.46% (59.66%–71.08%) | Of the 13 contrasts, six are primary (P1–P6) and seven secondary (S1–S7); none is adjusted for multiplicity. 95% intervals from 2,000 hierarchical paired bootstrap draws over people and training seeds (seed 20261007), conditional on the fixed person folds and three seeds; the five fixed rotations are averaged first. CCA has no training and a single seed axis; its only contrast is S7 (−67.97 pp). Held-out scores choose among 8 candidates and seen scores among 32, so the two are different tasks; m8 is the matched comparison for a held-out score, and there the FiLM head with the magnitude reference reached 42.86%. The contrasts against the plain spectrum (P1, P2, S3, S4) compare two feature recipes as well as the ramp: the complex and plain spectra differ in normalisation, size and capacity. The delay range is drawn from each window’s training label. The FiLM and matching heads start from different initial scores, so head comparisons are not initialisation-matched. BETA assigns each target its own stimulus phase, so phase carries real label information. The ramp acts on already-windowed coefficients, keeps magnitudes and the phase between channels, rotates the phases of the harmonics together, and approximates a time shift. All 10 randomized-label checks passed (120 fits); a training-only shortcut through the label-dependent delay remains possible. The 13 paired contrasts Each contrast is a paired difference in percentage points: the cell under its code, trained with the ramp, minus the arm in the comparison, which has the same head and query unless they are named. No ramp is the complex spectrum trained without the ramp; the plain spectrum has no ramp. | Contrast | Task | Comparison | Role | Estimate | 95% interval | | --- | --- | --- | --- | --- | --- | | `P1` — FiLM head · reference spectrum | Held-out | ramp minus plain spectrum | primary | −8.45 pp | −10.14 to −6.69 | | `P2` — matching head · reference spectrum | Held-out | ramp minus plain spectrum (FiLM head · reference spectrum) | primary | −26.07 pp | −28.84 to −23.19 | | `P3` — FiLM head · question number | Seen, balanced | ramp minus no ramp | primary | −46.38 pp | −51.06 to −41.71 | | `P4` — matching head · question number | Seen, balanced | ramp minus no ramp | primary | −52.98 pp | −57.96 to −48.11 | | `P5` — FiLM head · numeric code | Held-out | ramp minus no ramp | primary | +14.96 pp | +12.10 to +17.63 | | `P6` — matching head · numeric code | Held-out | ramp minus no ramp | primary | +2.09 pp | +0.88 to +3.29 | | `S1` — FiLM head · reference spectrum | Seen, balanced | ramp minus no ramp | secondary | −32.26 pp | −36.12 to −28.45 | | `S2` — matching head · reference spectrum | Seen, balanced | ramp minus no ramp | secondary | −50.94 pp | −56.07 to −45.96 | | `S3` — FiLM head · numeric code | Held-out | ramp minus plain spectrum | secondary | −13.37 pp | −15.60 to −11.30 | | `S4` — matching head · numeric code | Held-out | ramp minus plain spectrum (FiLM head · numeric code) | secondary | −29.04 pp | −32.22 to −25.73 | | `S5` — FiLM head · reference spectrum | Held-out | ramp minus no ramp | secondary | +21.88 pp | +18.78 to +25.00 | | `S6` — matching head · reference spectrum | Held-out | ramp minus no ramp | secondary | +6.41 pp | +5.40 to +7.34 | | `S7` — matching head · reference spectrum | Held-out | ramp minus CCA (no training) | secondary | −67.97 pp | −72.07 to −63.76 | Descriptive read-offs: generalisation cost, nn8 and u3 Descriptive read-offs with 95% intervals, defined under methods and limits. A dash marks a score that is not defined. | Head · query | Generalisation cost | nn8 | u3 | | --- | --- | --- | --- | | Complex spectrum, trained with the ramp | | | | | FiLM head · reference spectrum — `FS-REF` | −12.30 pp (−13.71 to −10.95) | 31.62% (28.79%–34.41%) | 25.09% (24.07%–26.22%) | | FiLM head · numeric code — `FS-NUM` | −7.25 pp (−8.62 to −5.97) | 28.42% (25.75%–30.96%) | 14.37% (13.62%–15.17%) | | FiLM head · question number — `FS-ID` | — | 13.74% (13.13%–14.36%) | — | | matching head · reference spectrum — `MATCH-REF` | +0.61 pp (−0.11 to +1.32) | 13.05% (12.49%–13.59%) | 24.20% (23.51%–24.92%) | | matching head · numeric code — `MATCH-NUM` | +0.05 pp (−0.64 to +0.73) | 12.50% (11.93%–13.10%) | 12.79% (12.28%–13.29%) | | matching head · question number — `MATCH-ID` | — | 13.46% (12.76%–14.17%) | — | | Complex spectrum, no ramp (matched controls) | | | | | FiLM head · reference spectrum — `L0c-FS-REF` | −56.89 pp (−61.14 to −52.56) | 8.56% (7.66%–9.53%) | 25.46% (24.22%–26.73%) | | FiLM head · numeric code — `L0c-FS-NUM` | −37.57 pp (−41.13 to −33.99) | 13.41% (12.44%–14.41%) | 7.16% (6.63%–7.79%) | | FiLM head · question number — `L0c-FS-ID` | — | 7.40% (6.58%–8.35%) | — | | matching head · reference spectrum — `L0c-MATCH-REF` | −63.64 pp (−68.11 to −59.06) | 6.69% (5.88%–7.54%) | 23.99% (23.12%–24.83%) | | matching head · numeric code — `L0c-MATCH-NUM` | −39.70 pp (−43.47 to −36.12) | 10.68% (9.68%–11.79%) | 10.76% (9.85%–11.71%) | | matching head · question number — `L0c-MATCH-ID` | — | 6.78% (5.90%–7.69%) | — | | Plain spectrum, no ramp | | | | | FiLM head · reference spectrum — `L0-FS-REF` | −18.18 pp (−19.64 to −16.69) | 43.41% (40.04%–46.58%) | 19.61% (18.53%–20.63%) | | FiLM head · numeric code — `L0-FS-NUM` | −12.30 pp (−13.42 to −11.17) | 40.96% (37.65%–44.13%) | 20.57% (19.33%–21.78%) | | No training | | | | | CCA (no training) — `CCA` | 0.00 pp | 66.66% (61.33%–71.56%) | 81.70% (78.49%–84.57%) | The matching head’s cost is near zero because its held-out and m8 scores both sit at chance. Card 5 · the matching computation ## Where CCA’s lead comes from CCA reads the raw window with no training. Which parts of that computation carry its lead over the learned heads? On the same splits, CCA on the raw two-second window reached 80.90% on held-out frequencies (76.80%–85.13%, person bootstrap). Tapering the window with a Hann function changed it by −7.95 pp (−9.07 to −6.78), to 72.96%. A cropped-spectrum match with no training reached 73.20% on held-out frequencies (68.62%–78.13%), +0.24 pp from Hann-tapered CCA (−0.12 to +0.64). It works on the complex spectral coordinates that the earlier complex-spectrum heads read before their fitted scaling; asked with the reference with stimulus phase, those heads reached 24.97% (FiLM head) and 22.47% (matching head) on [card 2](https://bci.report/topics/new-ssvep-frequencies/#head-and-features), quoted here with no paired contrast. The match fits channel and reference weights for each window without labels. Averaging each channel’s energy in the reference span reached 55.21% (49.99%–60.90%). One ridge scorer trained on the source people over each channel’s energy fraction at each of the 3 harmonics reached 57.45% (53.20%–62.14%): +2.23 pp over that average (+0.13 to +4.37) and −23.46 pp from CCA (−26.64 to −20.45). On the seen 32-way task it was −0.82 pp from the same average (−2.86 to +1.18). Card 5 on BETA: the seven prespecified contrasts, in percentage points, each with its 95% interval. | Contrast | Held-out 8-way | Seen 32-way balanced | Note | | --- | --- | --- | --- | | Hann-tapered CCA (centred) − CCA (no training) | −7.95 pp (95% interval −9.07 to −6.78) | −9.98 pp (95% interval −11.10 to −8.84) | — | | Hann-tapered CCA: uncentred − centred | 0.00 pp (95% interval 0.00 to 0.00) | 0.00 pp (95% interval 0.00 to 0.00) | Identical results from different definitions. | | Cropped-spectrum match (no training) − Hann-tapered CCA (uncentred) | +0.24 pp (95% interval −0.12 to +0.64) | +0.19 pp (95% interval −0.22 to +0.64) | Unresolved: the interval spans zero on both tasks. | | Cropped-spectrum match (no training) − channel-energy average (joint reference span) | +17.98 pp (95% interval +14.94 to +21.28) | +19.94 pp (95% interval +17.06 to +23.18) | This lead bundles three changes: cross-channel mixing, fitting the reference coefficients, and aggregation. | | Channel-energy average: per-harmonic spans − joint reference span | 0.00 pp (95% interval 0.00 to 0.00) | 0.00 pp (95% interval 0.00 to 0.00) | Identical results from different definitions. | | Shared ridge scorer − channel-energy average (per-harmonic spans) | +2.23 pp (95% interval +0.13 to +4.37) | −0.82 pp (95% interval −2.86 to +1.18) | Above the average on held-out frequencies; on seen frequencies the interval spans zero. | | Shared ridge scorer − CCA (no training) | −23.46 pp (95% interval −26.64 to −20.45) | −30.54 pp (95% interval −34.76 to −26.67) | — | Every scorer on both tasks Card 5 on BETA: accuracy of the seven scorers on held-out and seen frequencies, with 95% intervals. | Name on this page | Export code | Held-out 8-way | Seen 32-way balanced | | --- | --- | --- | --- | | CCA (no training) | `CCA-RAW` | 80.90% (95% interval 76.80%–85.13%) | 65.46% (95% interval 59.94%–71.40%) | | Hann-tapered CCA (centred) — no training | `CCA-HANN-C` | 72.96% (95% interval 68.36%–77.86%) | 55.48% (95% interval 50.01%–61.63%) | | Hann-tapered CCA (uncentred) — no training | `CCA-HANN-U` | 72.96% (95% interval 68.36%–77.86%) | 55.48% (95% interval 50.01%–61.63%) | | Cropped-spectrum match (no training) | `SUBSPACE-CROP` | 73.20% (95% interval 68.62%–78.13%) | 55.67% (95% interval 50.20%–61.78%) | | Channel-energy average (joint reference span) — no training | `CHANNEL-JOINT` | 55.21% (95% interval 49.99%–60.90%) | 35.74% (95% interval 30.77%–41.13%) | | Channel-energy average (per-harmonic spans) — no training | `CHANNEL-HARMONIC` | 55.21% (95% interval 49.99%–60.90%) | 35.74% (95% interval 30.77%–41.13%) | | Shared ridge scorer — ridge fitted on the source people | `SHARED-RIDGE` | 57.45% (95% interval 53.20%–62.14%) | 34.92% (95% interval 30.83%–39.20%) | Every method is given all 40 candidate frequencies; held-out means left out of supervised fitting. The study has 6 training-free scorers and 35 ridge fits, and seven prespecified contrasts on each task. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); no multiplicity correction. The cropped coordinates match the inputs the earlier heads read on the 6 rows checked; the heads’ fitted scaling is invertible but changes the distances the match uses. Checked: CCA reconstructed on all 11,200 windows, the training-free scorers spot-checked on 6 rows, all 35 ridge fits refitted independently. Card 6 · a learned correction to CCA ## Learning on top of CCA If CCA is kept and a learned part only corrects its score, does accuracy on held-out frequencies rise? Keeping temporal CCA and adding a bounded learned correction to its score reached 86.15% on held-out frequencies (82.84%–89.35%), against 80.90% for CCA alone: +5.24 pp (+3.91 to +6.62). On the seen 32-way task it reached 73.19% against 65.46% (+7.73 pp, +6.10 to +9.46). The correction can reorder only candidates whose CCA correlations lie within 0.2 of each other. Of all held-out decisions, it turned 7.39% from wrong to right and 2.14% from right to wrong. Giving the energy head six canonical correlations from the cropped-spectrum match, which makes it the adapted head, added +17.51 pp on held-out frequencies (+15.46 to +19.70); that gain bundles several matching operations. The candidate frequencies still supply CCA’s sine and cosine references and the numeric code of each frequency. Filter-bank CCA and supervised spatial templates were not compared. The adapted head against CCA alone was not a planned contrast and has no interval. Each trained arm is the mean of three fixed training seeds. Six prespecified contrasts on each task; 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015), conditional on the fitted models and the three fixed training seeds; no multiplicity correction. All six arms, the six contrasts and the decision changes Card 6 on BETA: accuracy of the six arms on held-out and seen frequencies, with 95% intervals. | Name on this page | Export code | Held-out 8-way | Seen 32-way balanced | | --- | --- | --- | --- | | Energy head — trained on the source people and seen frequencies | `ENERGY-FS` | 63.05% (95% interval 58.97%–67.50%) | 43.34% (95% interval 39.09%–47.83%) | | Adapted head — trained on the source people and seen frequencies | `ADAPT-FS` | 80.56% (95% interval 76.98%–84.13%) | 64.06% (95% interval 59.56%–68.85%) | | Cropped match with learned correction — trained on the source people and seen frequencies | `CROP-RESIDUAL` | 79.24% (95% interval 75.24%–83.37%) | 63.35% (95% interval 58.39%–68.72%) | | CCA with learned correction — trained on the source people and seen frequencies | `RAW-RESIDUAL` | 86.15% (95% interval 82.84%–89.35%) | 73.19% (95% interval 68.48%–77.99%) | | Cropped-spectrum match (no training) | `CCA-CROP` | 73.20% (95% interval 68.62%–78.13%) | 55.67% (95% interval 50.20%–61.78%) | | CCA (no training) | `CCA-RAW` | 80.90% (95% interval 76.80%–85.13%) | 65.46% (95% interval 59.94%–71.40%) | Card 6 on BETA: the six prespecified contrasts, in percentage points, each with its 95% interval. | Contrast | Held-out 8-way | Seen 32-way balanced | Note | | --- | --- | --- | --- | | Adapted head − energy head | +17.51 pp (95% interval +15.46 to +19.70) | +20.72 pp (95% interval +18.54 to +23.01) | — | | Adapted head − cropped-spectrum match (no training) | +7.36 pp (95% interval +5.33 to +9.32) | +8.38 pp (95% interval +5.91 to +10.82) | — | | Cropped match with learned correction − adapted head | −1.32 pp (95% interval −2.46 to −0.23) | −0.71 pp (95% interval −1.95 to +0.54) | On held-out frequencies the correction on the cropped match is below the adapted head. | | Cropped match with learned correction − cropped-spectrum match (no training) | +6.04 pp (95% interval +4.90 to +7.17) | +7.67 pp (95% interval +6.26 to +9.13) | — | | CCA with learned correction − CCA (no training) | +5.24 pp (95% interval +3.91 to +6.62) | +7.73 pp (95% interval +6.10 to +9.46) | — | | CCA with learned correction − cropped match with learned correction | +6.91 pp (95% interval +5.60 to +8.16) | +9.84 pp (95% interval +8.60 to +11.04) | The two corrections start from different baselines: CCA on the raw window, and the cropped match. | Card 6 on BETA: how the correction changed CCA’s decisions, as shares of all decisions on each task, averaged over three seeds, then five rotations, then 70 people. | Decision change | Held-out 8-way | Seen 32-way balanced | What it counts | | --- | --- | --- | --- | | Wrong to right | 7.39% | 10.66% | CCA alone chose a wrong frequency; with the correction the decision was right. | | Right to wrong | 2.14% | 2.93% | CCA alone chose the right frequency; with the correction the decision was wrong. | | Wrong to a different wrong | 2.59% | 7.63% | Both decisions wrong, on different frequencies. | | Unchanged, right | 78.76% | 62.53% | Both chose the right frequency. | | Unchanged, wrong | 9.12% | 16.25% | Both chose the same wrong frequency. | | Changed in total | 12.12% | 21.22% | The first three rows together. | Decisions whose two largest CCA correlations differed by more than 0.2 changed 0 times, out of 11,403 held-out and 6,012 seen prediction instances (counted over the 3 seeds and 5 rotations). Card 7 · a second dataset ## The same correction on Wearable-102 Run unchanged on a second dataset, does the learned correction still change CCA’s score? Wearable-102: 102 people, 12 targets from 9.25 to 14.75 Hz, 12,240 two-second windows per sensor, dry and wet headbands, recordings as stored; 105 unchanged BETA models. The documentation cannot rule out that some people also appear in BETA. The 105 models of [card 6](https://bci.report/topics/new-ssvep-frequencies/#cca-correction) (7 participant folds × 5 frequency rotations × 3 seeds), trained on BETA, were run unchanged on Wearable-102. On the same windows the models averaged 65.45% on dry electrodes and 81.04% on wet, and matched CCA 65.38% and 81.65%. Paired with matched CCA, the models changed the score by +0.07 pp dry (−0.71 to +0.91) and −0.61 pp wet (−1.48 to +0.23); both intervals include zero, and no equivalence margin was set. The models’ score is CCA plus a bounded learned term, so the change from matched CCA is what the learned term adds. Card 7: accuracy among the 12 Wearable-102 targets, by sensor, with 95% intervals. | Wearable-102 sensor | CCA with learned correction | Matched CCA | Change from matched CCA | Cropped-spectrum check (Wearable-102), diagnostic | | --- | --- | --- | --- | --- | | Wearable-102, dry — primary · 12,240 windows | 65.45% (95% interval 60.82%–70.08%) | 65.38% (95% interval 60.73%–70.06%) | +0.07 pp (95% interval −0.71 to +0.91) | 57.78% (95% interval 52.79%–62.89%) | | Wearable-102, wet — primary · 12,240 windows | 81.04% (95% interval 77.14%–84.49%) | 81.65% (95% interval 77.60%–85.14%) | −0.61 pp (95% interval −1.48 to +0.23) | 73.01% (95% interval 68.46%–77.18%) | | Wearable-102, both sensors — secondary · 24,480 windows | 73.24% (95% interval 69.40%–76.88%) | 73.51% (95% interval 69.71%–77.22%) | −0.27 pp (95% interval −0.95 to +0.40) | 65.40% (95% interval 60.94%–69.69%) | Matched CCA here uses sine and cosine references up to the third harmonic, as the models’ own CCA term does, and reached 65.38% dry and 81.65% wet; [the dry-versus-wet page](https://bci.report/topics/dry-vs-wet/)’s CCA uses references up to the fourth harmonic on the same windows and reached 65.47% and 81.56%. A cropped-spectrum check (Wearable-102) reached 57.78% dry and 73.01% wet; it is shown as a diagnostic only. **Limits.** Wearable-102’s stored amplitude unit is undocumented and the correction reads amplitude spectra, so these are as-stored figures. Its windows start 0.14 s after stimulus onset and BETA’s at onset, and hardware, reference, people and frequency grid change at the same time, so no single factor is isolated. Both methods receive the 12 candidate frequencies. Each person’s accuracy is computed for each of the 105 models and averaged over models, then over people, with no ensemble, refit or model choice. This task on other recordings stands apart from the BETA results of cards 1 to 6. Dry and wet are the two primary contrasts and both sensors the secondary one. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015), conditional on the fixed 105 source models; no multiplicity correction. Card 8 · a shared learned scorer ## A shared scorer on the cross-spectral matrix One set of learned weights, shared by every candidate frequency, scores each candidate’s channel-by-channel matrices at the fundamental and two harmonics. How far does that get? The cross-spectral ridge, one ridge readout over those matrices, reached 68.03% (63.39%–73.05%) on held-out frequencies against 57.45% ([card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) for the shared ridge scorer: +10.58 pp (+9.05 to +12.28), the first primary contrast. The cross-spectral MLP reached 72.00% (67.59%–76.65%) against CCA’s 80.90% ([card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)): −8.90 pp (−11.20 to −6.69), the second primary contrast; it did not pass its prespecified non-inferiority test against CCA. The whitened cross-spectral ridge reached 79.99% (76.20%–83.79%): −0.91 pp from CCA on held-out frequencies and −3.35 pp on seen ones. Each is the difference of two levels, with no planned contrast and no interval, so neither shows equivalence or non-inferiority. On the seen 32-way task every new arm is below CCA’s 65.46% ([card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)). Every set-up on both tasks, and the contrasts with their roles Card 8: accuracy on held-out frequencies and balanced accuracy on seen ones. The last four rows are card 5’s scores, which carry their intervals there. | Set-up | Training | Held-out 8-way | Seen 32-way, balanced | | --- | --- | --- | --- | | cross-spectral ridge — `XS-RIDGE` | ridge fitted on the source people | 68.03% (95% interval 63.39%–73.05%) | 48.56% (95% interval 43.56%–53.99%) | | whitened cross-spectral ridge — `XS-RIDGE-W` | ridge fitted on the source people | 79.99% (95% interval 76.20%–83.79%) | 62.10% (95% interval 57.28%–67.01%) | | cross-spectral MLP — `XS-MLP` | MLP trained on the source people, 3 seeds averaged | 72.00% (95% interval 67.59%–76.65%) | 50.08% (95% interval 45.26%–55.19%) | | cross-spectral ridge with quadrature — `XS-RIDGE-Q` | ridge fitted on the source people | 68.45% (95% interval 63.98%–73.26%) | 47.90% (95% interval 43.00%–53.36%) | | CCA (no training) — `CCA-RAW` | no training | 80.90% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 65.46% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | | cropped-spectrum match (no training) — `SUBSPACE-CROP` | no training | 73.20% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 55.67% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | | channel-energy average — `CHANNEL-HARMONIC` | no training | 55.21% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 35.74% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | | shared ridge scorer — `SHARED-RIDGE` | ridge fitted on the source people | 57.45% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 34.92% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | Card 8: paired differences, each with its own interval. | Contrast | Task | Role | Difference | | --- | --- | --- | --- | | cross-spectral ridge − shared ridge scorer | BETA held-out | primary | +10.58 pp (95% interval +9.05 to +12.28) | | cross-spectral MLP − CCA (no training) | BETA held-out | primary | −8.90 pp (95% interval −11.20 to −6.69) | | cross-spectral ridge − CCA (no training) | BETA held-out | secondary | −12.88 pp (95% interval −15.55 to −10.22) | | whitened cross-spectral ridge − cross-spectral ridge | BETA held-out | secondary | +11.96 pp (95% interval +9.63 to +14.30) | | cross-spectral MLP − cross-spectral ridge | BETA held-out | secondary | +3.97 pp (95% interval +2.86 to +5.10) | | cross-spectral ridge with quadrature − cross-spectral ridge | BETA held-out | secondary | +0.42 pp (95% interval −0.10 to +0.89) | | cross-spectral ridge − shared ridge scorer | BETA seen 32-way | secondary | +13.64 pp (95% interval +11.56 to +16.05) | | whitened cross-spectral ridge − shared ridge scorer | BETA seen 32-way | secondary | +27.19 pp (95% interval +24.00 to +30.59) | | cross-spectral MLP − shared ridge scorer | BETA seen 32-way | secondary | +15.16 pp (95% interval +13.02 to +17.60) | | cross-spectral ridge with quadrature − shared ridge scorer | BETA seen 32-way | secondary | +12.98 pp (95% interval +10.87 to +15.38) | Margins set before the run: equivalence within ±2 pp; non-inferiority to CCA, for the MLP only, a lower bound above −2 pp. The scorer is not a fixed quadratic form of the raw EEG: its diagonals are logged and its normalisation depends on the trial. The MLP differs from the ridge in nonlinearity, loss, regularisation, optimiser and capacity at once, and its seeds are averaged, not ensembled. Whitening changes the features and what the scorer is given at the same time. The cross-spectral arms keep the Hann-tapered, cropped spectral coordinates while CCA reads the raw window, so a contrast against CCA changes the representation as well. The matrices hold within-harmonic terms only, not every second-order interaction across harmonics. The quadrature arm minus the plain ridge is +0.42 pp (−0.10 to +0.89): the interval includes zero, so the direction is unresolved, which does not show quadrature information is useless to every model. The gain over the shared ridge scorer does not explain every earlier head’s failure. The cropped-spectrum match is context, not a ceiling for learned scorers. CCA, the shared ridge scorer, the cropped-spectrum match and the channel-energy average are [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)’s scores, reused unchanged. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015). Checks: 328 fits (210 main); CCA’s identity reproduced over 448,000 pairs; the release audit recomputed 7,000 metric entries without repeating the supervised fits. Card 9 · eigenvalue pooling ## Eigenvalues of the harmonic matrices, learned or summed A head on the eigenvalues of each harmonic’s matrix, learned or simply summed: what does learning add? A ridge head on the eigenvalues of the raw harmonic matrices, the eigenvalue head, reached 82.97% (79.09%–86.82%) on held-out frequencies against CCA’s 80.90% ([card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)): +2.07 pp (+1.09 to +3.12), the first primary contrast, above zero and inside the equivalence and non-inferiority margins the protocol set in advance (±5 pp). The harmonic sum of the same statistics, with no training, reached 82.90% (78.90%–86.93%); the head minus the sum is +0.07 pp (−0.45 to +0.62), unresolved. The other two primary contrasts are negative. The coherence head reached 64.35% (60.04%–69.13%) against the cross-spectral ridge’s 68.03% ([card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)): −3.68 pp (−4.92 to −2.46). The spectral eigenvalue head reached 74.92% (70.53%–79.44%) against the whitened cross-spectral ridge’s 79.99% ([card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)): −5.07 pp (−6.61 to −3.59). [Card 6](https://bci.report/topics/new-ssvep-frequencies/#cca-correction) listed a simpler learned head on the CCA features among the things it did not run. The eigenvalue head and the joint eigenvalue head are such heads; they read the squared canonical correlations between the raw window and the harmonic references, not card 6’s inputs. Every set-up on both tasks, and the contrasts with their roles Card 9: accuracy on held-out frequencies and balanced accuracy on seen ones, a separate task. Comparators from cards 5 and 8 carry their intervals there. | Set-up | Training | Held-out 8-way | Seen 32-way, balanced | | --- | --- | --- | --- | | coherence head — `COH-ABS` | ridge fitted on the source people | 64.35% (95% interval 60.04%–69.13%) | 40.29% (95% interval 36.24%–44.70%) | | signed coherence head — `COH-SIGNED` | ridge fitted on the source people | 66.68% (95% interval 62.29%–71.49%) | 42.53% (95% interval 38.24%–47.17%) | | spectral eigenvalue head — `EIG-SPEC` | ridge fitted on the source people | 74.92% (95% interval 70.53%–79.44%) | 57.31% (95% interval 51.91%–62.82%) | | eigenvalue head — `EIG-RAW` | ridge fitted on the source people | 82.97% (95% interval 79.09%–86.82%) | 70.00% (95% interval 64.72%–75.43%) | | joint eigenvalue head — `EIG-JOINT` | ridge fitted on the source people | 82.29% (95% interval 78.21%–86.39%) | 69.25% (95% interval 63.80%–74.98%) | | joint CCA — `JOINT-TOP` | no training | 80.90% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 65.46% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | | harmonic sum (no training) — `RAW-HARM-SUM` | no training | 82.90% (95% interval 78.90%–86.93%) | 69.53% (95% interval 64.02%–75.07%) | | spectral harmonic sum (no training) — `SPEC-HARM-SUM` | no training | 74.39% (95% interval 69.96%–79.06%) | 56.08% (95% interval 50.84%–61.57%) | | shared ridge scorer — `SHARED-RIDGE` | ridge fitted on the source people | 57.45% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 34.92% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | | CCA (no training) — `CCA-RAW` | no training | 80.90% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 65.46% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | | cross-spectral ridge — `XS-RIDGE` | ridge fitted on the source people | 68.03% (interval on [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)) | 48.56% (interval on [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)) | | whitened cross-spectral ridge — `XS-RIDGE-W` | ridge fitted on the source people | 79.99% (interval on [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)) | 62.10% (interval on [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)) | | cross-spectral ridge with quadrature — `XS-RIDGE-Q` | ridge fitted on the source people | 68.45% (interval on [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)) | 47.90% (interval on [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral)) | Card 9: paired differences, each with its own interval. | Contrast | Task | Role | Difference | | --- | --- | --- | --- | | coherence head − cross-spectral ridge | BETA held-out | primary | −3.68 pp (95% interval −4.92 to −2.46) | | spectral eigenvalue head − whitened cross-spectral ridge | BETA held-out | primary | −5.07 pp (95% interval −6.61 to −3.59) | | eigenvalue head − CCA (no training) | BETA held-out | primary | +2.07 pp (95% interval +1.09 to +3.12) | | signed coherence head − coherence head | BETA held-out | secondary | +2.33 pp (95% interval +1.62 to +3.10) | | signed coherence head − cross-spectral ridge | BETA held-out | secondary | −1.35 pp (95% interval −2.61 to −0.18) | | eigenvalue head − harmonic sum (no training) | BETA held-out | secondary | +0.07 pp (95% interval −0.45 to +0.62) | | spectral eigenvalue head − spectral harmonic sum (no training) | BETA held-out | secondary | +0.53 pp (95% interval −0.00022 to +1.07) | | joint eigenvalue head − joint CCA | BETA held-out | secondary | +1.38 pp (95% interval +0.56 to +2.18) | | coherence head − shared ridge scorer | BETA seen 32-way | secondary | +5.37 pp (95% interval +4.18 to +6.62) | | signed coherence head − shared ridge scorer | BETA seen 32-way | secondary | +7.61 pp (95% interval +6.14 to +9.16) | | spectral eigenvalue head − shared ridge scorer | BETA seen 32-way | secondary | +22.39 pp (95% interval +19.05 to +25.92) | | eigenvalue head − shared ridge scorer | BETA seen 32-way | secondary | +35.08 pp (95% interval +31.47 to +38.93) | | joint eigenvalue head − shared ridge scorer | BETA seen 32-way | secondary | +34.33 pp (95% interval +30.62 to +38.25) | The eigenvalue head and the joint eigenvalue head are given statistics derived from CCA for each trial, so their contrasts against CCA and joint CCA do not isolate what learning adds. The coherence arms use normalised projection cosines, not a conventional (Welch) coherence estimate, so they say nothing about coherence in general. The coherence head against the signed one isolates sign removal; against the cross-spectral ridge it changes normalisation and sign handling together. The spectral eigenvalue head against the whitened ridge also changes the feature dimension and the effective regularisation, and it differs from the raw eigenvalue head in more than whitening, so the raw head’s result does not validate the spectral one. The harmonic sum against joint CCA changes pooling, spatial weights and overlapping subspaces together, and eigenvalues taken harmonic by harmonic discard how the harmonic subspaces relate. The spectral head minus the spectral harmonic sum is +0.53 pp (−0.00022 to +1.07): its lower bound is printed to five decimals because at two it would read as zero. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); conditional on the fixed training and fold procedure. Checks: 361 fit operations (175 main, 175 independent refits); the eigenvalue identity reproduces CCA’s squared correlation, with identical decisions, over 448,000 pairs; the zero-energy rule gated 0 of 37,632,000 pairs. The comparators equal the scores of [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism) and [card 8](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral). Card 10 · harmonic pooling ## Pooling CCA’s per-harmonic statistics CCA gives one squared correlation per harmonic. Does the rule that pools them matter, and does a learned rule do better than a sum? ### BETA, held-out frequencies All three primary contrasts: the top-three ridge, learned on each harmonic’s leading correlation, minus the harmonic sum +0.30 pp (−0.12 to +0.78), unresolved; the eigenvalue head minus the top-three ridge −0.23 pp (−0.57 to +0.11), unresolved; the all-component sum, which sums all six components rather than each harmonic’s leading one, minus the harmonic sum −0.87 pp (−1.42 to −0.37). Secondary: the harmonic max minus the harmonic sum −7.72 pp (−9.12 to −6.46); the joint all-component sum minus CCA +1.12 pp (+0.26 to +1.94); the joint eigenvalue head minus the joint all-component sum +0.27 pp (−0.03 to +0.55). ### Wearable-102, dry and wet The same 12,240 windows per sensor as [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer), 102 people, 12 candidates in their native order, exact ties going to the first. Each learned arm’s 35 BETA heads run unchanged, their scores averaged with equal weights, which collapses them to one affine function; 0 fits on Wearable-102. The primary contrast, the eigenvalue head minus the harmonic sum: −0.36 pp (−1.06 to +0.31) dry and +0.36 pp (−0.24 to +0.94) wet, both unresolved. Every set-up on BETA and on Wearable-102, and all the contrasts Card 10, BETA: accuracy on held-out frequencies and balanced accuracy on seen ones (descriptive). Levels from cards 5 and 9 carry their intervals there. | Set-up | Training | Held-out 8-way | Seen 32-way, balanced | | --- | --- | --- | --- | | harmonic max (no training) — `RAW-HARM-MAX` | no training | 75.18% (95% interval 70.73%–80.03%) | 53.66% (95% interval 48.24%–59.84%) | | all-component sum (no training) — `RAW-HARM-ALLSUM` | no training | 82.03% (95% interval 77.95%–86.12%) | 68.89% (95% interval 63.36%–74.61%) | | joint all-component sum (no training) — `JOINT-ALLSUM` | no training | 82.02% (95% interval 77.95%–86.11%) | 68.86% (95% interval 63.33%–74.56%) | | top-three ridge — `RAW-LEAD3-RIDGE` | ridge fitted on the source people | 83.21% (95% interval 79.34%–87.07%) | 69.33% (95% interval 64.05%–74.69%) | | eigenvalue head — `EIG-RAW` | ridge fitted on the source people | 82.97% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 70.00% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | | joint eigenvalue head — `EIG-JOINT` | ridge fitted on the source people | 82.29% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 69.25% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | | harmonic sum (no training) — `RAW-HARM-SUM` | no training | 82.90% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 69.53% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | | CCA (no training) — `CCA-RAW` | no training | 80.90% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 65.46% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | Card 10, Wearable-102: accuracy among its targets, by sensor. Matched CCA is card 7’s level. | Set-up | Training | Wearable-102 dry | Wearable-102 wet | | --- | --- | --- | --- | | eigenvalue head — `EIG-RAW` | BETA heads, run unchanged | 65.68% (95% interval 60.87%–70.62%) | 83.13% (95% interval 79.83%–86.50%) | | joint eigenvalue head — `EIG-JOINT` | BETA heads, run unchanged | 66.79% (95% interval 62.29%–71.46%) | 83.45% (95% interval 80.11%–86.79%) | | top-three ridge — `RAW-LEAD3-RIDGE` | BETA heads, run unchanged | 64.51% (95% interval 59.67%–69.42%) | 82.26% (95% interval 78.84%–85.70%) | | matched CCA — `CCA-RAW` | no training | 65.38% (interval on [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer)) | 81.65% (interval on [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer)) | | harmonic sum (no training) — `RAW-HARM-SUM` | no training | 66.04% (95% interval 61.36%–70.87%) | 82.77% (95% interval 79.29%–86.17%) | | harmonic max (no training) — `RAW-HARM-MAX` | no training | 61.00% (95% interval 56.31%–65.98%) | 77.75% (95% interval 73.77%–81.59%) | | all-component sum (no training) — `RAW-HARM-ALLSUM` | no training | 66.83% (95% interval 62.25%–71.67%) | 83.54% (95% interval 80.32%–86.79%) | | joint all-component sum (no training) — `JOINT-ALLSUM` | no training | 66.79% (95% interval 62.22%–71.62%) | 83.48% (95% interval 80.24%–86.77%) | Card 10: paired differences, each with its own interval. | Contrast | Task | Role | Difference | | --- | --- | --- | --- | | top-three ridge − harmonic sum (no training) | BETA held-out | primary | +0.30 pp (95% interval −0.12 to +0.78) | | eigenvalue head − top-three ridge | BETA held-out | primary | −0.23 pp (95% interval −0.57 to +0.11) | | all-component sum (no training) − harmonic sum (no training) | BETA held-out | primary | −0.87 pp (95% interval −1.42 to −0.37) | | harmonic max (no training) − harmonic sum (no training) | BETA held-out | secondary | −7.72 pp (95% interval −9.12 to −6.46) | | joint all-component sum (no training) − CCA (no training) | BETA held-out | secondary | +1.12 pp (95% interval +0.26 to +1.94) | | joint eigenvalue head − joint all-component sum (no training) | BETA held-out | secondary | +0.27 pp (95% interval −0.03 to +0.55) | | eigenvalue head − harmonic sum (no training) | Wearable-102 dry | primary | −0.36 pp (95% interval −1.06 to +0.31) | | eigenvalue head − matched CCA | Wearable-102 dry | secondary | +0.30 pp (95% interval −0.73 to +1.27) | | top-three ridge − harmonic sum (no training) | Wearable-102 dry | secondary | −1.53 pp (95% interval −2.19 to −0.87) | | eigenvalue head − top-three ridge | Wearable-102 dry | secondary | +1.17 pp (95% interval +0.67 to +1.67) | | joint eigenvalue head − joint all-component sum (no training) | Wearable-102 dry | secondary | 0.00 pp (95% interval −0.34 to +0.31) | | eigenvalue head − harmonic sum (no training) | Wearable-102 wet | primary | +0.36 pp (95% interval −0.24 to +0.94) | | eigenvalue head − matched CCA | Wearable-102 wet | secondary | +1.48 pp (95% interval +0.58 to +2.39) | | top-three ridge − harmonic sum (no training) | Wearable-102 wet | secondary | −0.51 pp (95% interval −1.08 to +0.04) | | eigenvalue head − top-three ridge | Wearable-102 wet | secondary | +0.87 pp (95% interval +0.47 to +1.27) | | joint eigenvalue head − joint all-component sum (no training) | Wearable-102 wet | secondary | −0.03 pp (95% interval −0.30 to +0.24) | The eigenvalue features already carry CCA’s statistics for each trial. These contrasts compare pooling rules only: they do not show that second components are irrelevant in general, and the top-three result does not establish a learned advantage. No equivalence margin was registered, so an interval that includes zero leaves the direction unresolved. On Wearable-102 the average of the BETA heads is not a compact model trained on its own, and its levels are not comparable with BETA’s tasks; the candidates are supplied, so nothing here finds an unknown frequency, and the dataset’s units, windows and possible participant overlap are as [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer) states them. BETA: 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); paired across arms, with each person’s five rotations kept together. Wearable-102: 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261020); conditional on the fixed 35 source heads of each learned arm; paired across methods and both sensors, one set of draws. Card 11 · practical settings ## Filter-bank CCA, shorter windows and fewer electrodes Filter-bank CCA, shorter windows, and four electrodes instead of eight: how does each change held-out accuracy? BETA, held-out and seen tasks. Windows of 0.5 s, 1.0 s and 2.0 s cut from the authors’ 3–100 Hz preprocessing and cropped before any new filtering; eight recorded electrodes (PO5, PO3, POZ, PO4, PO6, O1, OZ, O2) or four of them (POZ, O1, OZ, O2). The raw-family methods filter 6–80 Hz and use 3 harmonics; filter-bank CCA filters each cropped window itself. Filter-bank CCA, which needs no training, reached 90.21% (87.65%–92.62%) on held-out frequencies with 2.0 s windows and eight electrodes. Against the harmonic sum’s 82.90% ([card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) that is +7.31 pp (+5.15 to +9.47), the primary contrast; against CCA’s 80.90% ([card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) it is +9.31 pp (+6.84 to +11.71). The eigenvalue head minus the harmonic sum is a primary contrast in each of the six cells. One is above zero, at 2.0 s with four electrodes: +0.87 pp (+0.08 to +1.70); the other five intervals include zero. Going from 2.0 s to 0.5 s windows changed filter-bank CCA by −55.39 pp (−57.46 to −53.28) with four electrodes and by −57.72 pp (−59.64 to −55.78) with eight. With 1.0 s windows, four electrodes instead of eight changed filter-bank CCA by −5.48 pp (−7.13 to −3.84) and the four raw-family methods by between −2.21 pp (−3.80 to −0.73) and −1.73 pp (−3.55 to −0.03), every interval below zero. With 0.5 s windows the harmonic sum, the top-three ridge and the eigenvalue head have higher level means with four electrodes than with eight, so eight electrodes do not always help; no paired interval was registered for those pairs. Card 11’s recipe is filter-bank CCA (seven Chebyshev sub-bands, five harmonics), our SciPy realisation of the published recipe. [Route 3’s page](https://bci.report/topics/questions-in-language/#secondary) reports another filter-bank recipe on the same task, S11, with five Butterworth sub-bands. [Card 6](https://bci.report/topics/new-ssvep-frequencies/#cca-correction) listed filter-bank CCA among the things it did not run; this card runs it on the authors’ preprocessed window, cropped before the filter bank. Every cell of the grid, on both tasks, and all the contrasts Card 11: accuracy on held-out frequencies, cells in the order fixed before the run. Levels shared with cards 5, 9 and 10 carry their intervals there. | Window · electrodes | CCA (no training) | harmonic sum (no training) | top-three ridge | eigenvalue head | filter-bank CCA (seven sub-bands) | | --- | --- | --- | --- | --- | --- | | 0.5 s · 4 electrodes | 23.99% (95% interval 22.08%–26.10%) | 24.54% (95% interval 22.69%–26.43%) | 24.80% (95% interval 23.01%–26.68%) | 25.08% (95% interval 23.20%–27.15%) | 30.40% (95% interval 28.14%–32.89%) | | 0.5 s · 8 electrodes | 24.33% (95% interval 22.35%–26.49%) | 24.09% (95% interval 22.35%–26.09%) | 24.20% (95% interval 22.47%–26.22%) | 24.42% (95% interval 22.69%–26.53%) | 32.49% (95% interval 30.00%–35.17%) | | 1.0 s · 4 electrodes | 52.78% (95% interval 47.86%–57.90%) | 54.59% (95% interval 49.78%–59.59%) | 55.39% (95% interval 50.70%–60.09%) | 55.46% (95% interval 50.78%–60.20%) | 63.77% (95% interval 59.51%–68.13%) | | 1.0 s · 8 electrodes | 54.99% (95% interval 50.38%–60.21%) | 56.80% (95% interval 52.11%–61.77%) | 57.13% (95% interval 52.54%–61.89%) | 57.38% (95% interval 52.90%–62.13%) | 69.25% (95% interval 65.13%–73.59%) | | 2.0 s · 4 electrodes | 77.21% (95% interval 72.67%–81.67%) | 78.92% (95% interval 74.39%–83.19%) | 79.54% (95% interval 75.36%–83.53%) | 79.79% (95% interval 75.67%–83.69%) | 85.79% (95% interval 82.49%–88.92%) | | 2.0 s · 8 electrodes | 80.90% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 82.90% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 83.21% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | 82.97% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 90.21% (95% interval 87.65%–92.62%) | Card 11: balanced accuracy on seen frequencies, a separate task, descriptive. | Window · electrodes | CCA (no training) | harmonic sum (no training) | top-three ridge | eigenvalue head | filter-bank CCA (seven sub-bands) | | --- | --- | --- | --- | --- | --- | | 0.5 s · 4 electrodes | 6.67% (95% interval 5.87%–7.52%) | 6.87% (95% interval 6.19%–7.60%) | 6.83% (95% interval 6.07%–7.63%) | 6.94% (95% interval 6.23%–7.67%) | 10.52% (95% interval 9.45%–11.71%) | | 0.5 s · 8 electrodes | 6.75% (95% interval 5.97%–7.63%) | 6.35% (95% interval 5.70%–7.11%) | 6.25% (95% interval 5.59%–6.98%) | 6.47% (95% interval 5.84%–7.23%) | 11.47% (95% interval 10.30%–12.86%) | | 1.0 s · 4 electrodes | 26.02% (95% interval 22.33%–29.93%) | 28.05% (95% interval 24.45%–31.92%) | 27.90% (95% interval 24.57%–31.52%) | 28.56% (95% interval 25.15%–32.17%) | 40.56% (95% interval 36.64%–44.94%) | | 1.0 s · 8 electrodes | 28.50% (95% interval 24.68%–32.87%) | 30.15% (95% interval 26.61%–34.27%) | 30.14% (95% interval 26.66%–34.14%) | 30.77% (95% interval 27.27%–34.66%) | 47.75% (95% interval 43.34%–52.60%) | | 2.0 s · 4 electrodes | 60.17% (95% interval 54.54%–66.11%) | 63.10% (95% interval 57.52%–68.82%) | 63.46% (95% interval 58.02%–68.80%) | 64.44% (95% interval 58.96%–69.60%) | 74.23% (95% interval 69.68%–78.83%) | | 2.0 s · 8 electrodes | 65.46% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 69.53% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 69.33% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | 70.00% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 80.76% (95% interval 76.94%–84.39%) | Card 11: paired differences on held-out frequencies, each with its own interval and its own role in this study. | Contrast | Group | Role | Difference | | --- | --- | --- | --- | | filter-bank CCA (seven sub-bands) − harmonic sum (no training), 2.0 s · 8 electrodes | method | primary | +7.31 pp (95% interval +5.15 to +9.47) | | eigenvalue head − harmonic sum (no training), 0.5 s · 4 electrodes | method | primary | +0.54 pp (95% interval −0.23 to +1.35) | | eigenvalue head − harmonic sum (no training), 0.5 s · 8 electrodes | method | primary | +0.33 pp (95% interval −0.56 to +1.19) | | eigenvalue head − harmonic sum (no training), 1.0 s · 4 electrodes | method | primary | +0.87 pp (95% interval −0.15 to +1.79) | | eigenvalue head − harmonic sum (no training), 1.0 s · 8 electrodes | method | primary | +0.57 pp (95% interval −0.29 to +1.36) | | eigenvalue head − harmonic sum (no training), 2.0 s · 4 electrodes | method | primary | +0.87 pp (95% interval +0.08 to +1.70) | | eigenvalue head − harmonic sum (no training), 2.0 s · 8 electrodes | method | primary | +0.07 pp (95% interval −0.45 to +0.62) | | top-three ridge − harmonic sum (no training), 0.5 s · 4 electrodes | method | secondary | +0.27 pp (95% interval −0.42 to +0.90) | | eigenvalue head − top-three ridge, 0.5 s · 4 electrodes | method | secondary | +0.28 pp (95% interval −0.19 to +0.79) | | filter-bank CCA (seven sub-bands) − CCA (no training), 0.5 s · 4 electrodes | method | secondary | +6.41 pp (95% interval +5.04 to +7.90) | | top-three ridge − harmonic sum (no training), 0.5 s · 8 electrodes | method | secondary | +0.11 pp (95% interval −0.49 to +0.71) | | eigenvalue head − top-three ridge, 0.5 s · 8 electrodes | method | secondary | +0.22 pp (95% interval −0.41 to +0.83) | | filter-bank CCA (seven sub-bands) − CCA (no training), 0.5 s · 8 electrodes | method | secondary | +8.16 pp (95% interval +6.60 to +9.77) | | top-three ridge − harmonic sum (no training), 1.0 s · 4 electrodes | method | secondary | +0.80 pp (95% interval −0.20 to +1.67) | | eigenvalue head − top-three ridge, 1.0 s · 4 electrodes | method | secondary | +0.06 pp (95% interval −0.37 to +0.52) | | filter-bank CCA (seven sub-bands) − CCA (no training), 1.0 s · 4 electrodes | method | secondary | +10.99 pp (95% interval +8.30 to +13.60) | | top-three ridge − harmonic sum (no training), 1.0 s · 8 electrodes | method | secondary | +0.32 pp (95% interval −0.38 to +0.99) | | eigenvalue head − top-three ridge, 1.0 s · 8 electrodes | method | secondary | +0.25 pp (95% interval −0.26 to +0.79) | | filter-bank CCA (seven sub-bands) − CCA (no training), 1.0 s · 8 electrodes | method | secondary | +14.26 pp (95% interval +11.67 to +16.75) | | top-three ridge − harmonic sum (no training), 2.0 s · 4 electrodes | method | secondary | +0.62 pp (95% interval −0.07 to +1.31) | | eigenvalue head − top-three ridge, 2.0 s · 4 electrodes | method | secondary | +0.26 pp (95% interval −0.13 to +0.66) | | filter-bank CCA (seven sub-bands) − CCA (no training), 2.0 s · 4 electrodes | method | secondary | +8.58 pp (95% interval +6.15 to +10.87) | | top-three ridge − harmonic sum (no training), 2.0 s · 8 electrodes | method | secondary | +0.30 pp (95% interval −0.12 to +0.78) | | eigenvalue head − top-three ridge, 2.0 s · 8 electrodes | method | secondary | −0.23 pp (95% interval −0.57 to +0.11) | | filter-bank CCA (seven sub-bands) − CCA (no training), 2.0 s · 8 electrodes | method | secondary | +9.31 pp (95% interval +6.84 to +11.71) | | eigenvalue head − filter-bank CCA (seven sub-bands), 2.0 s · 8 electrodes | method | secondary | −7.24 pp (95% interval −9.30 to −5.18) | | CCA (no training): 0.5 s · 4 electrodes − 2.0 s · 4 electrodes | window | secondary | −53.22 pp (95% interval −56.38 to −49.87) | | CCA (no training): 0.5 s · 8 electrodes − 2.0 s · 8 electrodes | window | secondary | −56.57 pp (95% interval −59.86 to −53.27) | | harmonic sum (no training): 0.5 s · 4 electrodes − 2.0 s · 4 electrodes | window | secondary | −54.38 pp (95% interval −57.51 to −51.08) | | harmonic sum (no training): 0.5 s · 8 electrodes − 2.0 s · 8 electrodes | window | secondary | −58.81 pp (95% interval −61.86 to −55.64) | | top-three ridge: 0.5 s · 4 electrodes − 2.0 s · 4 electrodes | window | secondary | −54.73 pp (95% interval −57.70 to −51.44) | | top-three ridge: 0.5 s · 8 electrodes − 2.0 s · 8 electrodes | window | secondary | −59.01 pp (95% interval −61.96 to −55.92) | | eigenvalue head: 0.5 s · 4 electrodes − 2.0 s · 4 electrodes | window | secondary | −54.71 pp (95% interval −57.54 to −51.66) | | eigenvalue head: 0.5 s · 8 electrodes − 2.0 s · 8 electrodes | window | secondary | −58.55 pp (95% interval −61.54 to −55.43) | | filter-bank CCA (seven sub-bands): 0.5 s · 4 electrodes − 2.0 s · 4 electrodes | window | secondary | −55.39 pp (95% interval −57.46 to −53.28) | | filter-bank CCA (seven sub-bands): 0.5 s · 8 electrodes − 2.0 s · 8 electrodes | window | secondary | −57.72 pp (95% interval −59.64 to −55.78) | | CCA (no training): 1.0 s · 4 electrodes − 1.0 s · 8 electrodes | electrodes | secondary | −2.21 pp (95% interval −3.87 to −0.73) | | harmonic sum (no training): 1.0 s · 4 electrodes − 1.0 s · 8 electrodes | electrodes | secondary | −2.21 pp (95% interval −3.80 to −0.73) | | top-three ridge: 1.0 s · 4 electrodes − 1.0 s · 8 electrodes | electrodes | secondary | −1.73 pp (95% interval −3.55 to −0.03) | | eigenvalue head: 1.0 s · 4 electrodes − 1.0 s · 8 electrodes | electrodes | secondary | −1.92 pp (95% interval −3.70 to −0.21) | | filter-bank CCA (seven sub-bands): 1.0 s · 4 electrodes − 1.0 s · 8 electrodes | electrodes | secondary | −5.48 pp (95% interval −7.13 to −3.84) | At 2.0 s with eight electrodes the grid reproduces the levels and contrasts of [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling) and [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)’s eigenvalue head minus harmonic sum; two contrasts that are primary on card 10 are secondary here, each keeping its own study’s role. CPU component timings Warm, single-thread wall-clock medians of 3 calls after 1 warm-up call, with the fastest and slowest call, for the cell at 2.0 s with eight electrodes, scoring 40 candidates with one fixed released head. Filter-bank CCA took 7.45 ms for a single trial and 2.82 ms per trial in a batch of 8. Each component is timed on its own and never summed: the shared raw CCA geometry serves the four raw-family methods and is timed once, the readouts exclude feature extraction, and the decision step and the average over heads were not timed. A batch figure divides the batch’s time by its size and is not the latency of a lone trial. The cell was not chosen for speed, and nothing here compares or recommends hardware. Card 11: CPU timings in milliseconds — components; not sensor-to-decision or online latency. | Component | Code | One trial (ms) | Batch of 8, per trial (ms) | | --- | --- | --- | --- | | crop, validation and copy | `crop` | 0.0497 (range 0.0416–0.0534) | 0.00769 (range 0.00761–0.00874) | | raw 6–80 Hz filter | `raw_filter` | 1.23 (range 0.948–1.28) | 0.147 (range 0.122–0.181) | | shared raw CCA geometry | `raw_geometry` | 1.39 (range 1.23–1.39) | 0.315 (range 0.306–0.405) | | filter-bank CCA: filter, geometry and scores | `fbcca` | 7.45 (range 6.99–8.02) | 2.82 (range 2.81–2.92) | | top-three ridge readout on cached statistics | `top3_readout` | 0.0303 (range 0.0263–0.0311) | 0.00690 (range 0.00682–0.00697) | | eigenvalue head readout on cached statistics | `eigenvalue_readout` | 0.0595 (range 0.0594–0.0623) | 0.00798 (range 0.00755–0.00803) | Filter-bank CCA against the raw-family methods compares complete pipelines: its sub-bands, harmonics and weighting change together. A window length is how much signal the decoder sees, not acquisition-to-decision latency, and every cell inherits the authors’ preprocessing. Four electrodes are a recorded subset of the eight, not a four-electrode headset. The candidate frequencies’ references are given at test time, so nothing here supports a 40-way or online spelling claim, and the filter-bank result is no evidence about foundation models. The recipe is a SciPy realisation, not a bitwise replication of the authors’ MATLAB code. No equivalence margin was registered, and the seen-task levels are descriptive. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); paired across methods, cells and rotations within each sampled person. Card 12 · the harmonic sum’s gain ## Where the harmonic sum’s gain over joint CCA comes from Joint CCA and the harmonic sum read the same three harmonic references. Which step gives the sum its gain? On BETA’s held-out frequencies the path runs from joint CCA, 80.90% ([card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)), to the common-direction sum, 80.94% (76.79%–84.90%), to the harmonic sum, 82.90% ([card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)). Joint CCA to the common-direction sum: +0.04 pp (−0.10 to +0.19); the common-direction sum to the harmonic sum: +1.96 pp (+1.06 to +2.96). Both are 99.44% intervals of a Bonferroni family over nine comparisons, the primary family. The true candidate’s relaxation gap G exceeds the distractors’ mean in all three cohorts, in squared-correlation units, not percentages: BETA 0.0617 (0.0519 to 0.0732), Wearable-102 dry 0.0232 (0.0186 to 0.0286) and wet 0.0435 (0.0355 to 0.0525). On Wearable-102 the common-direction sum to the harmonic sum is +0.66 pp (−0.25 to +1.58) dry, unresolved, and +1.17 pp (+0.35 to +2.01) wet; joint CCA to the common-direction sum is 0.00 pp (−0.16 to +0.16) dry and −0.05 pp (−0.19 to +0.08) wet. No step from joint CCA to the common-direction sum is resolved in any cohort. Separate harmonic references are not orthogonal over a finite window, so joint CCA squared is not in general the largest eigenvalue of the summed matrices; the coupling residual R = B − J is reported beside G = S − B, and R takes either sign. The three cohorts, the gap statistics, the allocation and the decision changes Card 12: accuracy of the four scores, none trained. Joint CCA and the harmonic sum carry their intervals on the cards named. | Score | BETA held-out | Wearable-102 dry | Wearable-102 wet | | --- | --- | --- | --- | | joint CCA — `J` | 80.90% (interval on [card 5](https://bci.report/topics/new-ssvep-frequencies/#cca-mechanism)) | 65.38% (interval on [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer)) | 81.65% (interval on [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer)) | | common-direction sum — `B` | 80.94% (95% interval 76.79%–84.90%) | 65.38% (95% interval 60.56%–70.04%) | 81.60% (95% interval 77.78%–85.18%) | | harmonic sum (no training) — `S` | 82.90% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 66.04% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | 82.77% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | | joint CCA plus gap (diagnostic score) — `JplusG` | 82.88% (95% interval 78.84%–86.71%) | 66.08% (95% interval 61.23%–70.73%) | 82.75% (95% interval 79.17%–86.13%) | Card 12: paired differences. The primary family’s intervals are Bonferroni-adjusted; the secondary intervals are descriptive. | Contrast | Task | Role | Difference | | --- | --- | --- | --- | | common-direction sum − joint CCA | BETA held-out | primary (family) | +0.04 pp (99.44% interval −0.10 to +0.19) | | harmonic sum (no training) − common-direction sum | BETA held-out | primary (family) | +1.96 pp (99.44% interval +1.06 to +2.96) | | harmonic sum (no training) − joint CCA | BETA held-out | secondary | +2.00 pp (95% interval +1.32 to +2.74) | | joint CCA plus gap (diagnostic score) − joint CCA | BETA held-out | secondary | +1.97 pp (95% interval +1.29 to +2.71) | | harmonic sum (no training) − joint CCA plus gap (diagnostic score) | BETA held-out | secondary | +0.03 pp (95% interval −0.04 to +0.09) | | common-direction sum − joint CCA | Wearable-102 dry | primary (family) | 0.00 pp (99.44% interval −0.16 to +0.16) | | harmonic sum (no training) − common-direction sum | Wearable-102 dry | primary (family) | +0.66 pp (99.44% interval −0.25 to +1.58) | | harmonic sum (no training) − joint CCA | Wearable-102 dry | secondary | +0.66 pp (95% interval −0.01 to +1.34) | | joint CCA plus gap (diagnostic score) − joint CCA | Wearable-102 dry | secondary | +0.70 pp (95% interval +0.05 to +1.36) | | harmonic sum (no training) − joint CCA plus gap (diagnostic score) | Wearable-102 dry | secondary | −0.04 pp (95% interval −0.16 to +0.08) | | common-direction sum − joint CCA | Wearable-102 wet | primary (family) | −0.05 pp (99.44% interval −0.19 to +0.08) | | harmonic sum (no training) − common-direction sum | Wearable-102 wet | primary (family) | +1.17 pp (99.44% interval +0.35 to +2.01) | | harmonic sum (no training) − joint CCA | Wearable-102 wet | secondary | +1.12 pp (95% interval +0.56 to +1.71) | | joint CCA plus gap (diagnostic score) − joint CCA | Wearable-102 wet | secondary | +1.09 pp (95% interval +0.54 to +1.67) | | harmonic sum (no training) − joint CCA plus gap (diagnostic score) | Wearable-102 wet | secondary | +0.02 pp (95% interval −0.07 to +0.12) | Card 12: gap statistics in squared-correlation units, and the pairwise win fraction of the true candidate’s G over each distractor’s, half credit for exact ties (not an accuracy). | Quantity | BETA held-out | Wearable-102 dry | Wearable-102 wet | | --- | --- | --- | --- | | G, true candidate | 0.121 (95% interval 0.114 to 0.129) | 0.0624 (95% interval 0.0588 to 0.0662) | 0.0881 (95% interval 0.0821 to 0.0944) | | G, mean of the distractors | 0.0596 (95% interval 0.0588 to 0.0604) | 0.0392 (95% interval 0.0386 to 0.0398) | 0.0446 (95% interval 0.0437 to 0.0454) | | G, true minus mean distractor | 0.0617 (99.44% interval 0.0519 to 0.0732) | 0.0232 (99.44% interval 0.0186 to 0.0286) | 0.0435 (99.44% interval 0.0355 to 0.0525) | | R, true minus mean distractor | −0.0000389 (95% interval −0.0000714 to −0.00000482) | −0.0000681 (95% interval −0.000108 to −0.0000289) | −0.000161 (95% interval −0.000226 to −0.000100) | | Pairwise win fraction (not an accuracy) | 0.857 (95% interval 0.837 to 0.876) | 0.747 (95% interval 0.728 to 0.766) | 0.831 (95% interval 0.814 to 0.848) | Card 12: descriptive accounting of the harmonic sum’s gain over joint CCA, the average of the two prespecified orders; not a cause. | Quantity | BETA held-out | Wearable-102 dry | Wearable-102 wet | | --- | --- | --- | --- | | Share through G | +1.97 pp (95% interval +1.30 to +2.69) | +0.68 pp (95% interval +0.04 to +1.33) | +1.13 pp (95% interval +0.57 to +1.72) | | Share through R | +0.03 pp (95% interval −0.03 to +0.09) | −0.02 pp (95% interval −0.11 to +0.07) | −0.01 pp (95% interval −0.09 to +0.06) | Card 12: decisions that change between scores, as shares of all 11,200 BETA held-out decisions. The rescued and harmed sets of different steps can overlap. | Step | Wrong to right | Right to wrong | Any change of choice | Right in both | Wrong in both | | --- | --- | --- | --- | --- | --- | | joint CCA → common-direction sum | 0.17% | 0.13% | 0.44% | 80.77% | 18.93% | | common-direction sum → harmonic sum (no training) | 4.00% | 2.04% | 8.47% | 78.90% | 15.06% | | joint CCA → harmonic sum (no training) | 4.01% | 2.01% | 8.44% | 78.89% | 15.09% | | joint CCA → joint CCA plus gap (diagnostic score) | 4.00% | 2.03% | 8.42% | 78.88% | 15.10% | | joint CCA plus gap (diagnostic score) → harmonic sum (no training) | 0.11% | 0.08% | 0.29% | 82.79% | 17.02% | Reference check, label-free and with no outcome subgroups: for 8 of 40 BETA candidates and 0 of 12 Wearable-102 candidates the cross-block norm between the harmonic reference blocks falls at or below the threshold, so those blocks are orthogonal to numerical precision; for the other candidates they are not. G is an association: a larger gap for the true candidate fits harmonic-specific spatial structure but does not show a cause, and the mean distractor is not the strongest competitor. Wet and dry gaps come from the same people and were not tested against each other. 95% percentile intervals from 20,000 paired person bootstrap draws (seeds 20261021 for BETA and 20261022 for Wearable-102); Wearable-102 dry and wet share one set of person draws; the primary family’s 9 comparisons use Bonferroni-adjusted 99.44% intervals. Card 13 · weights on the harmonics ## One decay rule on harmonics and on filter-bank bands The filter-bank recipe weights its sub-bands by a decaying rule. What does the same rule do on the three harmonics? One fixed weight rule, n^−1.25 + 0.25, the sub-band weights of the published filter-bank CCA recipe. On the three harmonics it lowered BETA held-out accuracy from 82.90% ([card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) to 79.91%: −2.99 pp (−3.77 to −2.29), the primary contrast. Its prespecified 90% interval, −3.64 to −2.38, lies wholly outside the ±0.5 pp margin, so the equivalence criterion is not met. On filter-bank CCA’s seven sub-bands the same rule raised accuracy from 85.01% to 90.21% ([card 11](https://bci.report/topics/new-ssvep-frequencies/#practical-settings)): +5.21 pp (+4.08 to +6.38). On the seen 32-way task the rule changed the harmonic sum by −5.72 pp (−6.64 to −4.84) and filter-bank CCA by +7.40 pp (+6.07 to +8.79). On Wearable-102, where only the harmonic construction ran, it changed accuracy by −1.28 pp (−2.05 to −0.54) dry and −1.36 pp (−1.91 to −0.79) wet. Every set-up on every task, and the contrasts Card 13: accuracy (balanced on the seen task). The export gives these levels without intervals; the harmonic sum’s and filter-bank CCA’s are on the cards named, and each contrast carries its own. | Set-up | Held-out 8-way | Seen 32-way, balanced | Wearable-102 dry | Wearable-102 wet | | --- | --- | --- | --- | --- | | harmonic sum (no training) — `H3-FLAT` | 82.90% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 69.53% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 66.04% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | 82.77% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | | harmonic sum with decay weights (no training) — `H3-DECAY` | 79.91% | 63.81% | 64.75% | 81.41% | | filter-bank CCA with flat sub-band weights (no training) — `FBCCA-FLAT` | 85.01% | 73.36% | — | — | | filter-bank CCA (seven sub-bands) — `FBCCA-DECAY` | 90.21% (interval on [card 11](https://bci.report/topics/new-ssvep-frequencies/#practical-settings)) | 80.76% (interval on [card 11](https://bci.report/topics/new-ssvep-frequencies/#practical-settings)) | — | — | Card 13: paired differences, each with its own 95% interval. | Contrast | Task | Role | Difference | | --- | --- | --- | --- | | harmonic sum with decay weights (no training) − harmonic sum (no training) | BETA held-out | primary | −2.99 pp (95% interval −3.77 to −2.29) | | harmonic sum with decay weights (no training) − harmonic sum (no training) | BETA seen 32-way | secondary | −5.72 pp (95% interval −6.64 to −4.84) | | harmonic sum with decay weights (no training) − harmonic sum (no training) | Wearable-102 dry | secondary | −1.28 pp (95% interval −2.05 to −0.54) | | harmonic sum with decay weights (no training) − harmonic sum (no training) | Wearable-102 wet | secondary | −1.36 pp (95% interval −1.91 to −0.79) | | filter-bank CCA (seven sub-bands) − filter-bank CCA with flat sub-band weights (no training) | BETA held-out | secondary | +5.21 pp (95% interval +4.08 to +6.38) | | filter-bank CCA (seven sub-bands) − filter-bank CCA with flat sub-band weights (no training) | BETA seen 32-way | secondary | +7.40 pp (95% interval +6.07 to +8.79) | Harmonic weights and sub-band weights are different operations, and each contrast changes weights within one construction only; applying the sub-band rule to harmonic numbers adapts the published recipe. The decayed filter-bank arm is [card 11](https://bci.report/topics/new-ssvep-frequencies/#practical-settings)’s filter-bank CCA. A single fixed rule says nothing about an optimal weighting. The margin test reads a percentile interval, so its equivalence inference is approximate, and the secondary intervals are descriptive. Wearable-102’s units and windows are as [card 7](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer) states them. 95% percentile intervals from 20,000 paired person bootstrap draws (seeds 2026102201 for BETA and 2026102202 for Wearable-102); Wearable-102 dry and wet share one set of person draws; the primary contrast’s margin test uses a 90% interval against a margin of ±0.5 pp. Card 14 · weights on the harmonics ## A rising weight on the harmonics Does weighting the three harmonics 1, 1.5, 1.9, instead of equally, raise held-out accuracy? On BETA’s held-out frequencies, rising weights on the three harmonics raised accuracy from the harmonic sum’s 82.90% ([card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) to 83.76% (79.91%–87.32%): +0.86 pp (+0.29 to +1.45). The single prespecified criterion, a paired 95% interval wholly above zero, is met. On the seen 32-way task the change was +0.49 pp (−0.22 to +1.21). On Wearable-102 the same weights changed accuracy from 66.04% ([card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) to 64.89% (59.86%–69.76%) dry, −1.14 pp (−1.86 to −0.46), and by −0.43 pp (−1.03 to +0.16) wet. The rising weights follow a coverage heuristic: filter-bank CCA’s sub-bands share one upper edge, so a higher harmonic falls in more of them. The exact profile divides the cumulative sum of [card 13](https://bci.report/topics/new-ssvep-frequencies/#fixed-decay)’s decay weights over the three harmonics by its first term; its rows are a rounding check, not an equivalence claim. The actual filter passbands are wider than their nominal edges, so a candidate’s harmonics can fall in more sub-bands than the nominal edges suggest; coverage counts and summed band weights are heuristics. No pre-stimulus alpha or noise was measured, and no weight was learned or searched. Every set-up on every task, and the contrasts Card 14: accuracy (balanced on the seen task). The harmonic sum carries its intervals on the cards named. | Set-up | Held-out 8-way | Seen 32-way, balanced | Wearable-102 dry | Wearable-102 wet | | --- | --- | --- | --- | --- | | harmonic sum (no training) — `flat` | 82.90% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 69.53% (interval on [card 9](https://bci.report/topics/new-ssvep-frequencies/#eigenvalue-pooling)) | 66.04% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | 82.77% (interval on [card 10](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling)) | | rising weights 1, 1.5, 1.9 (no training) — `literal_rise` | 83.76% (95% interval 79.91%–87.32%) | 70.02% (95% interval 64.66%–75.12%) | 64.89% (95% interval 59.86%–69.76%) | 82.34% (95% interval 78.77%–85.73%) | | rising weights, exact profile (no training) — `exact_rise` | 83.71% (95% interval 79.86%–87.29%) | 69.96% (95% interval 64.62%–75.05%) | 64.85% (95% interval 59.81%–69.71%) | 82.30% (95% interval 78.75%–85.69%) | Card 14: paired differences, each with its own interval; the exact-profile rows are a rounding check. | Contrast | Task | Role | Difference | | --- | --- | --- | --- | | rising weights 1, 1.5, 1.9 (no training) − harmonic sum (no training) | BETA held-out | primary | +0.86 pp (95% interval +0.29 to +1.45) | | rising weights, exact profile (no training) − harmonic sum (no training) | BETA held-out | rounding check | +0.81 pp (95% interval +0.22 to +1.43) | | rising weights, exact profile (no training) − rising weights 1, 1.5, 1.9 (no training) | BETA held-out | rounding check | −0.04 pp (95% interval −0.12 to +0.02) | | rising weights 1, 1.5, 1.9 (no training) − harmonic sum (no training) | BETA seen 32-way | secondary | +0.49 pp (95% interval −0.22 to +1.21) | | rising weights, exact profile (no training) − harmonic sum (no training) | BETA seen 32-way | rounding check | +0.44 pp (95% interval −0.32 to +1.21) | | rising weights, exact profile (no training) − rising weights 1, 1.5, 1.9 (no training) | BETA seen 32-way | rounding check | −0.05 pp (95% interval −0.16 to +0.05) | | rising weights 1, 1.5, 1.9 (no training) − harmonic sum (no training) | Wearable-102 dry | secondary | −1.14 pp (95% interval −1.86 to −0.46) | | rising weights, exact profile (no training) − harmonic sum (no training) | Wearable-102 dry | rounding check | −1.18 pp (95% interval −1.91 to −0.48) | | rising weights, exact profile (no training) − rising weights 1, 1.5, 1.9 (no training) | Wearable-102 dry | rounding check | −0.04 pp (95% interval −0.12 to +0.05) | | rising weights 1, 1.5, 1.9 (no training) − harmonic sum (no training) | Wearable-102 wet | secondary | −0.43 pp (95% interval −1.03 to +0.16) | | rising weights, exact profile (no training) − harmonic sum (no training) | Wearable-102 wet | rounding check | −0.47 pp (95% interval −1.09 to +0.15) | | rising weights, exact profile (no training) − rising weights 1, 1.5, 1.9 (no training) | Wearable-102 wet | rounding check | −0.03 pp (95% interval −0.11 to +0.03) | The dry interval lies below zero; the wet and seen intervals include zero. Apart from the single primary contrast, the intervals are descriptive. 95% percentile intervals from 20,000 paired person bootstrap draws (seeds 2026102501 for BETA and 2026102502 for Wearable-102); Wearable-102 dry and wet share one set of person draws. Methods & limits ## Methods and limits ### Definitions - **Plain spectrum (L0)**. The log-magnitude spectrum of each of the eight channels, kept channel by channel: it keeps which channel is which and discards phase. - **Complex spectrum (L0c)**. Each channel’s complex Fourier coefficients of the Hann-tapered window, real and imaginary parts, so phase is kept. It also differs from the plain spectrum in normalisation, size and the capacity of the head that reads it. - **Frozen CBraMod features (L1)**. The features of the frozen CBraMod encoder that route 3 used; no encoder weight was trained here. - **m8**. The same eight frequencies and windows, scored among eight candidates in the four other rotations, where those frequencies were seen in training: the matched comparison for a held-out score. - **Generalisation cost**. Held-out accuracy minus m8 for the same arm. For a scorer with no training it is zero by construction. - **nn8**. A neighbour read-off: for each held-out frequency, the arm’s scores for the adjacent seen frequency on either side are averaged (the one neighbour at the edge of the grid), and the choice is then made among the eight held-out candidates. It interpolates from the neighbours, beside the direct held-out score. - **u3**. A local diagnostic: for windows whose frequency is held out, the score for that frequency competes with its adjacent seen frequencies (two candidates at the edge of the grid). It uses the true target to choose its candidates, which makes it a test of local frequency discrimination; on frozen CBraMod features it also involves stimulus phase. - **Ties**. Candidate frequencies are in ascending order, and an exact score tie selects the lowest frequency. ### Intervals, card by card - **Cards 1 and 2**. 95% intervals from 2,000 hierarchical paired bootstrap draws over people and fold-shared seed indices (seed 20261007). - **Cards 3 and 4**. 95% intervals from 2,000 hierarchical paired bootstrap draws over people and training seeds (seed 20261007), conditional on the fixed person folds and three seeds; the five fixed rotations are averaged first. - **Card 5**. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015). - **Card 6**. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015), conditional on the fitted models and the three fixed training seeds. - **Card 7**. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015), conditional on the fixed 105 source models. - **Card 8**. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015). - **Card 9**. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); conditional on the fixed training and fold procedure. - **Card 10**. BETA: 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); paired across arms, with each person’s five rotations kept together; Wearable-102: 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261020); conditional on the fixed 35 source heads of each learned arm; paired across methods and both sensors, one set of draws. - **Card 11**. 95% percentile intervals from 2,000 paired person bootstrap draws (seed 20261015); paired across methods, cells and rotations within each sampled person. - **Card 12**. 95% percentile intervals from 20,000 paired person bootstrap draws (seeds 20261021 for BETA and 20261022 for Wearable-102); Wearable-102 dry and wet share one set of person draws; the primary family’s 9 comparisons use Bonferroni-adjusted 99.44% intervals. - **Card 13**. 95% percentile intervals from 20,000 paired person bootstrap draws (seeds 2026102201 for BETA and 2026102202 for Wearable-102); Wearable-102 dry and wet share one set of person draws; the primary contrast’s margin test uses a 90% interval against a margin of ±0.5 pp. - **Card 14**. 95% percentile intervals from 20,000 paired person bootstrap draws (seeds 2026102501 for BETA and 2026102502 for Wearable-102); Wearable-102 dry and wet share one set of person draws. Card 12’s primary family uses Bonferroni-adjusted 99.44% intervals over nine comparisons, and card 13’s equivalence test a 90% interval; no other interval is adjusted. Differences are paired and computed before rounding, so a printed difference can differ in its last digit from the difference of the printed levels. ### Engineering disclosures Card 2 scored the earlier question-number and numeric-code runs through a reviewed compatibility reader, because those files lacked a field listing their questions. The softmax run’s pre-freeze guard was changed to hash file contents. Neither change touched training, predictions or results. - Card 8, before scoring: a package’s metadata and its module version declaration did not match and were corrected before BETA was processed; the centring in a scalar identity diagnostic ran in float32 and was corrected to float64, with the tolerances unchanged. Neither changed the scientific protocol or used supervised accuracy. - Card 9, before execution: the frozen zero-channel-energy rule is now enforced where squaring the channel energy underflows, covered by a synthetic regression test; an initial synthetic test suite timed out and is not counted as a pass; a bounded rerun on synthetic data only passed all 20 of its tests; no real input had been processed before the edge correction. ### Not run - **Card 1**: a synthetic reference waveform passed through CBraMod; the matching (bilinear) head (run in card 2); a fundamental-only reference (run in card 2); a hyperparameter search; an independent replication on a new dataset. - **Card 3**: a control with an independent delay on each channel; the same delay with the numeric-code query (run in card 4); the matching head (run in card 4); cutting a new raw window at the delayed onset; delays at test time; features derived from CCA, or a learned correction to it (run in card 6); transfer to a new dataset. - **Card 4**: an extra matching-head arm whose delay ignores the label; a plain softmax trained with the same delay; a test-time check of how far scores change under the ramp; cutting a new raw window at the delayed onset; another dataset or model; sweeps over the learning rate or the delay range. - **Card 5**: a paired contrast with the earlier complex-spectrum heads. - **Card 6**: filter-bank CCA; supervised spatial templates; removing the trunk or the query; a simpler learned head on the CCA features. - **Card 7**: an equivalence margin; changing one factor at a time: units, window onset, hardware, people or the frequency grid; adapting or retraining on the target data. - **Card 8**: a learned cell on the raw time series; second-order terms across harmonics. - **Card 9**: a conventional averaged (Welch) coherence estimator; a neural-network head; an equivalence margin for the eigenvalue head against the harmonic sum. - **Card 10**: an equivalence margin; paired contrasts on the seen 32-way task; fitting, scaling or selecting heads on the target data; attributing a change to one factor: units, window, reference, hardware or cohort. - **Card 11**: a minimum practical margin; ablations of the filter bank, the harmonic count and the pooling; online performance or information transfer rate; a physical four-electrode headset. - **Card 12**: a learned head on the same statistics; eigenvector stability, source separation or a physiological intervention. - **Card 13**: other exponents, offsets or harmonic counts. - **Card 14**: a pre-stimulus alpha or noise measurement; a weight search or frequency-dependent normalisation. ### Not published here **Cards 1 to 7 (new-ssvep-frequencies-update.json)** - Per-person, per-fold, per-rotation, per-seed, per-window and per-frequency values of any kind, predictions, scores, features, fitted heads and the sealed run files: only cohort levels, paired contrasts with their intervals, transition shares and check counts are published. - The handoffs, protocols, freezes, receipts and reviews, and card 3’s canonical score file: each is pinned by SHA-256 in the release manifest and checked by the export; several carry private storage paths. - Card 6’s per-seed accuracies and the decision transitions of the cropped match with learned correction; the Wearable-102 contrast against the cropped-spectrum check. **Cards 8 to 14 (ssvep-followups-update.json)** - Per-person, per-fold, per-rotation, per-seed and per-window values of any kind, predictions, scores, features, fitted heads and the sealed run files: only cohort levels, paired contrasts with their intervals, gap statistics, decision-change shares, check counts and CPU component medians are published. - Per-frequency values of any kind, including the reference check’s candidate frequencies: the check is published as counts. - The cross-spectral MLP’s per-seed accuracies, and the canary and random-pairing accuracies of cards 8 and 9. - Card 13’s retrospective gap-to-gain ratio table and its dry reconstruction. - Decision counts over all trials: card 12’s decision changes on BETA held-out frequencies are published as shares of the 11,200 decisions; the Wearable-102 decision changes and the transitions of cards 13 and 14 are not published. - Card 12’s Delta means and the margins against the strongest joint-CCA distractor. - Run times of whole evaluations and audits, memory and process identifiers; the CPU-time readings of the component measurements, and every host, processor, system and runtime label. - Predictions from the comment thread on cards 8 to 14: each of these cards states only its prespecified criteria, and the thread’s prediction stays on card 4. - The commenter’s name. - The handoffs, protocols, freezes, receipts, reviews and approvals themselves: each is pinned by SHA-256 in the release manifest and checked by the export; several carry private storage paths. ### Credits, licences and consent ### BETA Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Cards 1 to 6 and 8 to 14, and the source of the models card 7 runs on Wearable-102. Attribution required. Participant consent described use for non-commercial scientific research, a narrower operational boundary than the copyright license; minors required parental consent. Cohort aggregates only; no participant records, EEG signals, or individual predictions. [Source ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ### BETA authors’ corrigendum (Description.pdf) The stimulus phases of the reference with stimulus phase (cards 2 and 3), converted to the cosine convention with zero latency. Read from the Hugging Face mirror Bingchuan/BETA (revision `d4290c02…`). [Description.pdf ↗](https://huggingface.co/datasets/Bingchuan/BETA/resolve/d4290c0200db8a104e0f557349dc49f90ba79506/Description.pdf) ### CBraMod The frozen CBraMod features of card 1 (L1), the ones route 3 used. No weights are distributed. Code: MIT · Weights: Apache-2.0 (official model card) [CBraMod’s page](https://bci.report/methods/cbramod/) · [Code ↗](https://github.com/wjq-learning/CBraMod) ### Wearable-102 · An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces. Sensors, 21(4), 1256. https://doi.org/10.3390/s21041256 Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 Cards 7, 10, 12, 13 and 14. The paper describes consent for noncommercial scientific research. Computed from the official Tsinghua BCI Lab author mirror snapshot acquired 2026-09-20; the mirror itself has no separate DOI or version number. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.13560281.v4) · [Readme ↗](https://bci.med.tsinghua.edu.cn/upload/zhufangkun/Readme.pdf) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ### FBCCA · Chen et al. 2015 Chen X, Wang Y, Gao S, Jung T-P, Gao X. Filter bank canonical correlation analysis for implementing a high-speed SSVEP-based brain–computer interface. Journal of Neural Engineering 12(4):046008, 2015. Cards 11 and 13: filter-bank CCA (seven Chebyshev sub-bands, five harmonics), our SciPy realisation of the published recipe. No code is distributed. [FBCCA’s page](https://bci.report/methods/fbcca/) · [Paper ↗](https://doi.org/10.1088/1741-2560/12/4/046008) Derived analyses by BCI Report, aggregates only; no raw data are redistributed. Data source: [new-ssvep-frequencies-update.json](https://bci.report/data/new-ssvep-frequencies-update.json) · schema bci-report-new-ssvep-frequencies-update-v1; [ssvep-followups-update.json](https://bci.report/data/ssvep-followups-update.json) · schema bci-report-ssvep-followups-update-v1. ## Related questions - Route 3’s own result on unseen frequencies, a head asked in words: [Can an EEG model answer questions asked in words?](https://bci.report/topics/questions-in-language/#unseen-frequency) - Wearable-102’s dry and wet headbands, with CCA and trained decoders: [Do dry EEG electrodes decode as well as wet ones?](https://bci.report/topics/dry-vs-wet/) A longer write-up of these follow-ups is in [Towards AI ↗](https://pub.towardsai.net/can-ai-improve-a-20-year-old-brain-decoder-a-little-and-not-everywhere-88397b8a69b7) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *What lets a decoder answer an SSVEP frequency it never trained on?* https://bci.report/topics/new-ssvep-frequencies/ Figures from releases [`ssvep-followups-update-20261028`](https://bci.report/releases/#ssvep-followups-update-20261028) (2026-10-28), [`new-ssvep-frequencies-update-20261020`](https://bci.report/releases/#new-ssvep-frequencies-update-20261020) (2026-10-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/new-ssvep-frequencies/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # How often does an EEG decoder fire when nobody is giving a command? Abstention · reliable decisions · Jev-style research plan On this site’s idle protocol, EEGNet made no false activation in 60 idle trials but detected 1 of 60 commands. On EEGMAT and BETA, route 1 of a Jev-style research plan measures when a decoder should decline to decide. - [Reviewed aggregate JSON ↓](https://bci.report/data/experiments.json) - Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds005342](https://bci.report/datasets/ds005342/); [YSU asynchronous SSVEP-BCI dataset](https://bci.report/datasets/ysu-async-ssvep/) · Methods: [EEGNet](https://bci.report/methods/eegnet/); [LaBraM](https://bci.report/methods/labram/); [CBraMod](https://bci.report/methods/cbramod/); [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/); [Deep4Net](https://bci.report/methods/deep4net/); [CSP+LDA](https://bci.report/methods/csp-lda/); [CCA](https://bci.report/methods/cca/) ## Short answer It depends on how often the decoder acts at all: one that never acts never fires by mistake, so its false activations mean something only beside the commands it caught. In a cue-gated idle test with **4** people, the quietest decoder, EEGNet, made **0 / 60** false activations in idle trials but detected **1 / 60** commands, and for **2** of those people its threshold rejected every command. In a separate test with **20** other people, a CCA decoder with a threshold fixed on an earlier pilot accepted **748 / 960** windows when a command was intended, and **49 / 240**, **93 / 240** and **107 / 480** in three states where none was. A threshold fitted instead on **96** of each person’s own windows accepted fewer non-control windows overall, and fewer in each state, but did not raise the share of commands both accepted and correct; the contrast cannot separate fitting the threshold to the person from simply having those labels. **0 / 60** EEGNet: idle false activations, 4 people ### No false activation, and 1 command in 60 Each threshold is chosen on calibration data only; for 2 of 4 people it rejected every command. Both numbers, every method ## Command detection against idle false activation 4 people; 60 idle and 60 command test trials in total; a command counts if it is detected within 3 seconds of its cue. Sorted by command detection. Counts over the whole test set, with the percentage beneath. “Always abstained” counts people whose calibrated threshold rejected every command. | Method | Training | Commands detected ≤3 s | Idle false activations | Always abstained | | --- | --- | --- | --- | --- | | ShallowFBCSPNet | Scratch · 20 epochs | 26 / 60 (43.3%) | 1 / 60 (1.7%) | 1 (of 4 people) | | CSP+LDA | Supervised fit | 14 / 60 (23.3%) | 0 / 60 (0.0%) | 1 (of 4 people) | | LaBraM | Frozen encoder + linear head | 11 / 60 (18.3%) | 2 / 60 (3.3%) | 0 (of 4 people) | | CBraMod | Frozen encoder + linear head | 11 / 60 (18.3%) | 1 / 60 (1.7%) | 1 (of 4 people) | | Deep4Net | Scratch · 20 epochs | 2 / 60 (3.3%) | 1 / 60 (1.7%) | 1 (of 4 people) | | EEGNet | Scratch · 20 epochs | 1 / 60 (1.7%) | 0 / 60 (0.0%) | 2 (of 4 people) | Explicit non-control · 20 people · rules fixed before scoring; global threshold from a 4-person pilot ## Does it accept a window when nobody intends a command? A second data set records the states between commands: looking at a central image with the flicker off, resting with the eyes on a white wall, and looking at the centre while the surrounding targets flicker. A fixed CCA decoder accepts or rejects each 1.5-second window as a command, under two rules fixed before any of these 20 people was scored: a global threshold fitted only on the calibration windows of the 4 pilot people below, and a personal threshold fitted on 96 labelled windows from each new person. Both rules score the same 96 held-out windows of each person. **+3.0 pp** Detection balanced accuracy, personal minus global threshold, same 20 people ### A personal threshold detected intent a little better — and produced no more correct commands Paired interval +0.5 to +5.7 pp. 10 of 20 people improved, 8 got worse and 2 were unchanged. The gain came from fewer false acceptances: coverage of intended commands barely moved, and the share of command windows both accepted and correct went from 68.1% to 67.6%. When a command was intended: 960 windows under each rule, 48 from each person. | Measure | Global threshold · fixed on the pilot | Personal threshold · own windows | | --- | --- | --- | | Detection balanced accuracy, command against no command | 76.0% (95% interval 71.6%–80.2%) | 79.0% (95% interval 74.2%–83.8%) | | Command windows accepted (coverage) | 748 / 960 (77.9%) | 746 / 960 (77.7%) | | Accepted and correct (end to end) | 654 / 960 (68.1%) | 649 / 960 (67.6%) | | Correct among accepted | 85.8% (mean of each person’s rate) | 85.1% (mean of each person’s rate) | Before any rejection the frequency was recognised in 769 / 960 command windows under both rules: the decoder is the same, and only the rejection rule differs. Detection went up; correct commands did not. When no command was intended: windows accepted by mistake in each state, under each rule. | State | Global threshold · fixed on the pilot | Personal threshold · own windows | | --- | --- | --- | | Central image, flicker off — NS1 | 49 / 240 (20.4%) | 32 / 240 (13.3%) | | Looking at a white wall, resting — NS2 | 93 / 240 (38.8%) | 79 / 240 (32.9%) | | Central image while the surrounding targets flicker — NS3 | 107 / 480 (22.3%) | 79 / 480 (16.5%) | | All three states, pooled — every non-control window | 249 / 960 (25.9%) | 190 / 960 (19.8%) | The detection score pools the three states as they were collected, 12, 12 and 24 windows a person, so the flickering-centre state counts twice; the last row is that pool. ### Limits of the non-control test ### Same dataset The 20 people come from the same release, lab and protocol as the pilot; this is not evidence from an independent cohort. ### 96 labels per person The personal rule uses 96 labelled windows from each person, half with a command and half without. The comparison cannot separate fitting the threshold to the person from simply having those labels. The source does not say when those windows were recorded relative to the test windows, so calibrating earlier and using later is untested here. ### Window rates Separately collected 1.5-second windows in a fixed mix of states. They give window rates, not false activations per hour of continuous use, latency, or how often people actually issue commands. ### No amplitude unit The source states no physical unit for its EEG values. The decoder uses correlations, which are unit-free. ### Bootstrap intervals Whole-person bootstrap over the 20, conditional on the 4 pilot people and on each realised calibration split. ### Development set: the 4-person pilot Scored with per-person thresholds fitted on each person’s own calibration trials. Its 4 people supplied the global threshold above and are not among the 20. Pooled window counts. When a command was intended · Rejecting hard windows raises accuracy among accepted windows while refusing commands the person meant. | State | Windows | | --- | --- | | Frequency recognised, before rejection | 152 / 192 (79.2%) | | Accepted (coverage) | 130 / 192 (67.7%) | | Accepted and correct (end to end) | 121 / 192 (63.0%) | | Correct among accepted | 121 / 130 (93.1%) | When no command was intended | State | Accepted by mistake | | --- | --- | | Central image, flicker off | 9 / 48 (18.8%) | | Looking at a white wall, resting | 20 / 48 (41.7%) | | Central image while the surrounding targets flicker | 11 / 96 (11.5%) | Offline, window by window. [Why four people limit what can be published (English)](https://bci.report/data-use/#small-cohorts) ### YSU asynchronous SSVEP-BCI dataset Jing Zhao, Qian Zhang, Xinrui Wang, Xueshuo Liu, Jiaxin Li, Fengjie Fan, Zhenhu Liang and Xiaoli Li · An EEG dataset for studying asynchronous steady-state visual evoked potential (SSVEP) based brain computer interfaces, Brain-Apparatus Communication 3(1) (2024), doi:10.1080/27706710.2024.2418650; data at doi:10.6084/m9.figshare.24906300.v3. Before the experiment, every participant signed written informed consent; the study was reviewed and approved by the ethics committee of Qinhuangdao First Hospital. This site uses all twenty-four people of the release, four as the development pilot and twenty as the extension, and publishes only cohort means, pooled window counts and how many people were helped or harmed. [Data paper ↗](https://doi.org/10.1080/27706710.2024.2418650) · [figshare release ↗](https://doi.org/10.6084/m9.figshare.24906300.v3) · [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) Data source: [extension-update.json](https://bci.report/data/extension-update.json) · schema bci-report-extension-update-v1. Data source: [reviewed aggregate JSON](https://bci.report/data/context-update.json) · schema bci-report-context-update-v1. Methods & limits ## Limits of the idle test ### 4 people At this size an average stays close to its parts: a published rate can be turned back into a count, and a count narrowed to one of two people. ### Three minutes of idle The test holds 180 seconds of idle, too little to estimate false activations per hour of unprompted rest. ### Cue-gated Detection restarts at every cue, so the decoder always knows a decision window has opened. A system that must notice intent on its own faces a harder problem than this protocol measures. ### No probability quality on this protocol These are thresholded decisions. Probability calibration is measured on EEGMAT and BETA, in the reliable-decisions results below. [The four-person arithmetic, in full (English)](https://bci.report/data-use/#small-cohorts) ### ds005342 · Seated motor imagery · cue-gated replay OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. The source README reports voluntary signed consent and approval by the ethics committee of Universidad Antonio Nariño. Only cohort aggregates appear here. [OpenNeuro dataset ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) · [Associated study ↗](https://doi.org/10.3389/fninf.2022.961089) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/experiments.json) · [this protocol as CSV](https://bci.report/data/idle-results.csv) · research-preview-20260920. Reliable decisions · route 1 of the plan below · EEGMAT and BETA ## When should a decoder decline to decide? Route 1 took the saved test scores of models this site already publishes — on EEGMAT, 36 people, mental arithmetic against rest, and on BETA, 70 people, a 40-target SSVEP speller — and asked which trials each decoder should have declined, under four rules fixed in advance. Full-coverage accuracies are the published ones: for the two LaBraM arms, the adaptation release’s seed-20260922 mean. ### Four rules, fixed before scoring - **Fixed threshold (F)**: accept a trial when its calibrated top-class probability is at least 0.8. Temperature scaling, fitted on calibration scores, keeps every predicted label. - **Coverage target (S-cov)**: accept the most confident trials, with the cut-off set from calibration quantiles so that 80% should be accepted. - **Certified risk (S-risk)**: selection with guaranteed risk (SGR, δ = 0.05) certifies a cut-off whose error should stay at or below a target r*, here half the calibration error. A fold where nothing can be certified accepts nothing. - **Learned reject option (L)**: a logistic selector on the classifier’s top two scores and their margin, plus a spectral description of the signal that does not depend on the model, trained to predict which trials the classifier gets wrong. Calibration scores never include a test person: in each outer fold the training people form four groups, and each group is scored by a model trained on the other three. Intervals are paired person-bootstrap 95% intervals over 2,000 resamples, with no multiplicity correction. A contrast is called a difference only when its interval excludes zero; differences of rates are in percentage points (pp). ### A fixed threshold is a different policy for each method The same cut-off — accept when the calibrated top-class probability is at least 0.8 — accepts very different shares of trials from different decoders, and from different people. BETA · 70 people · 11,200 trials, uniform over 40 targets. Methods in the core matrix’s order. | Method | Balanced accuracy, everything accepted | Accepted at confidence ≥ 0.8 | Error among accepted | People with nothing accepted | Coverage when the cut-off targets 80% | | --- | --- | --- | --- | --- | --- | | Standard CCA | 63.1% (the published value) | 3,176 / 11,200 (coverage 28.4% · 95% interval 22.9%–34.2%) | 4.3% (95% interval 2.7%–6.8%) | 0 (of 70 people) | 79.7% (95% interval 75.0%–84.0%) | | CBraMod | 33.7% (the published value) | 165 / 11,200 (coverage 1.5% · 95% interval 0.8%–2.3%) | 22.4% (95% interval 5.9%–45.1%) | 38 (of 70 people) | 76.4% (95% interval 72.8%–80.0%) | | EEGNet | 55.8% (the published value) | 1,863 / 11,200 (coverage 16.6% · 95% interval 12.5%–21.0%) | 5.5% (95% interval 2.8%–9.1%) | 10 (of 70 people) | 80.5% (95% interval 75.6%–85.1%) | BETA: paired differences in coverage under the fixed threshold, first method minus second | Pair | Difference | Reading | | --- | --- | --- | | Standard CCA − CBraMod | +26.9 pp (95% interval +21.7 to +32.4 pp) | difference resolved | | Standard CCA − EEGNet | +11.7 pp (95% interval +7.8 to +16.2 pp) | difference resolved | | CBraMod − EEGNet | −15.2 pp (95% interval −19.2 to −11.5 pp) | difference resolved | EEGMAT · 36 people · 2,160 windows, half rest and half arithmetic. Methods in the core matrix’s order. | Method | Balanced accuracy, everything accepted | Accepted at confidence ≥ 0.8 | Error among accepted | People with nothing accepted | Coverage when the cut-off targets 80% | | --- | --- | --- | --- | --- | --- | | Spectral ridge | 56.8% (the published value) | 3 / 2,160 (coverage 0.1% · 95% interval 0.0%–0.3%) | 0 / 3 (wrong, of those accepted) Unstable: fewer than 10 accepted, so no rate is read from it | 33 (of 36 people) | 80.6% (95% interval 78.9%–82.3%) | | EEGNet | 67.6% (the published value) | 126 / 2,160 (coverage 5.8% · 95% interval 2.3%–10.6%) | 42.1% (95% interval 13.0%–61.4%) | 19 (of 36 people) | 79.1% (95% interval 74.9%–83.2%) | | LaBraM · head only, encoder frozen | 56.9% (the published seed-20260922 mean, adaptation release) | 1 / 2,160 (coverage 0.05% · 95% interval 0.0%–0.1%) | 0 / 1 (wrong, of those accepted) Unstable: fewer than 10 accepted, so no rate is read from it | 35 (of 36 people) | 79.4% (95% interval 77.7%–81.1%) | | LaBraM · LoRA rank 4 + head | 64.4% (the published seed-20260922 mean, adaptation release) | 44 / 2,160 (coverage 2.0% · 95% interval 0.9%–3.5%) | 22.7% (95% interval 6.3%–45.5%) | 24 (of 36 people) | 81.5% (95% interval 79.1%–83.9%) | EEGMAT: paired differences in coverage under the fixed threshold, first method minus second | Pair | Difference | Reading | | --- | --- | --- | | Spectral ridge − EEGNet | −5.7 pp (95% interval −10.5 to −2.1 pp) | difference resolved | | Spectral ridge − LaBraM (head only) | +0.1 pp (95% interval 0.00 to +0.2 pp) | no difference resolved | | Spectral ridge − LaBraM (LoRA) | −1.9 pp (95% interval −3.3 to −0.8 pp) | difference resolved | | EEGNet − LaBraM (head only) | +5.8 pp (95% interval +2.2 to +10.6 pp) | difference resolved | | EEGNet − LaBraM (LoRA) | +3.8 pp (95% interval +0.4 to +8.2 pp) | difference resolved | | LaBraM (head only) − LaBraM (LoRA) | −2.0 pp (95% interval −3.4 to −0.9 pp) | difference resolved | #### Error at matched coverage On BETA the smallest coverage the fixed threshold reached was 165 of 11,200 trials (1.5%, CBraMod’s). Each method then accepts its own 165 most confident trials, so the threshold drops out of the comparison: BETA: error among each method’s 165 most confident trials, with paired differences. | Method | Error among its most confident | Reading | | --- | --- | --- | | Standard CCA | 0.6% (95% interval 0.0%–2.3%) | a descriptive level, no verdict | | CBraMod | 22.4% (95% interval 5.9%–45.1%) | a descriptive level, no verdict | | EEGNet | 8.5% (95% interval 0.0%–28.5%) | a descriptive level, no verdict | | Standard CCA − CBraMod | −21.8 pp (95% interval −44.6 to −5.1 pp) | difference resolved | | Standard CCA − EEGNet | −7.9 pp (95% interval −28.2 to +1.7 pp) | no difference resolved | | CBraMod − EEGNet | +13.9 pp (95% interval +5.3 to +22.2 pp) | difference resolved | On EEGMAT that common coverage is 1 window of 2,160: fewer than 10 accepted, so no verdict. ### A learned reject option did not beat the model’s own calibrated confidence Holding the classifier fixed, the learned reject option L — which also sees a spectral description of the signal — was set against the calibrated confidence S of the same classifier, on the same test trials. Two measures: the area under the risk–coverage curve (AURC; lower means the errors sit among the trials ranked least certain), and the error among the 80% of trials each ranks as most certain. L had a lower error than S at 80% for no method; where a difference was resolved, L erred more. By AURC, L ranked better only for CBraMod on BETA and worse for LaBraM (head only) on EEGMAT; every other contrast: no difference resolved. EEGMAT · 36 people. | Method | AURC · confidence S | AURC · learned L | L − S, AURC | Error at 80% · S | Error at 80% · L | L − S, error at 80% | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge | 0.386 (95% interval 0.348–0.428) | 0.399 (95% interval 0.359–0.440) | +0.013 (95% interval −0.006 to +0.029) no difference resolved | 41.4% (95% interval 38.9%–43.9%) | 41.9% (95% interval 39.2%–44.4%) | +0.5 pp (95% interval −1.2 to +1.7 pp) no difference resolved | | EEGNet | 0.293 (95% interval 0.211–0.368) | 0.309 (95% interval 0.238–0.379) | +0.016 (95% interval −0.014 to +0.050) no difference resolved | 29.5% (95% interval 23.9%–35.0%) | 31.8% (95% interval 27.0%–36.9%) | +2.3 pp (95% interval +0.6 to +4.1 pp) difference resolved | | LaBraM · head only, encoder frozen | 0.396 (95% interval 0.351–0.441) | 0.422 (95% interval 0.378–0.466) | +0.026 (95% interval +0.005 to +0.046) difference resolved | 42.9% (95% interval 39.0%–45.8%) | 43.6% (95% interval 40.2%–46.6%) | +0.6 pp (95% interval −0.5 to +2.7 pp) no difference resolved | | LaBraM · LoRA rank 4 + head | 0.308 (95% interval 0.245–0.373) | 0.323 (95% interval 0.267–0.379) | +0.015 (95% interval −0.011 to +0.041) no difference resolved | 33.2% (95% interval 28.6%–37.4%) | 35.1% (95% interval 30.7%–39.8%) | +1.9 pp (95% interval +0.6 to +3.5 pp) difference resolved | BETA · 70 people. | Method | AURC · confidence S | AURC · learned L | L − S, AURC | Error at 80% · S | Error at 80% · L | L − S, error at 80% | | --- | --- | --- | --- | --- | --- | --- | | Standard CCA | 0.145 (95% interval 0.105–0.194) | 0.146 (95% interval 0.107–0.196) | +0.002 (95% interval −0.004 to +0.008) no difference resolved | 27.2% (95% interval 20.8%–33.9%) | 27.4% (95% interval 21.1%–34.5%) | +0.3 pp (95% interval −0.3 to +1.0 pp) no difference resolved | | CBraMod | 0.497 (95% interval 0.430–0.566) | 0.473 (95% interval 0.413–0.538) | −0.024 (95% interval −0.039 to −0.007) difference resolved | 61.7% (95% interval 56.8%–66.5%) | 61.6% (95% interval 57.1%–66.4%) | −0.03 pp (95% interval −0.8 to +0.6 pp) no difference resolved | | EEGNet | 0.208 (95% interval 0.161–0.264) | 0.201 (95% interval 0.154–0.257) | −0.007 (95% interval −0.014 to +0.001) no difference resolved | 34.8% (95% interval 28.7%–41.4%) | 35.6% (95% interval 29.2%–42.4%) | +0.8 pp (95% interval +0.02 to +1.6 pp) difference resolved | #### What declining buys, under the calibrated confidence How much declining buys is read per protocol; the two are not compared: forty targets against two classes. On BETA, standard CCA’s error falls from 36.9% with everything accepted to 12.4% among its most certain half; on EEGMAT, EEGNet’s falls from 32.4% to 27.1%. Error among the trials accepted, by the share of test trials the calibrated confidence accepts. Point estimates. | Method | Everything accepted | Most certain 80% | Most certain 50% | | --- | --- | --- | --- | | EEGMAT | | | | | Spectral ridge | 43.2% | 41.4% | 39.5% | | EEGNet | 32.4% | 29.5% | 27.1% | | LaBraM · head only, encoder frozen | 43.1% | 42.9% | 40.7% | | LaBraM · LoRA rank 4 + head | 35.6% | 33.2% | 29.5% | | BETA | | | | | Standard CCA | 36.9% | 27.2% | 12.4% | | CBraMod | 66.3% | 61.7% | 53.2% | | EEGNet | 44.2% | 34.8% | 18.8% | ### A certified error level did not reliably hold on new people S-risk’s guarantee assumes calibration and test trials are exchangeable; across people they are not. On EEGMAT it certified no outer fold for any method, so nothing was accepted. On BETA it certified folds for every method, and in some certified folds the error on the new people exceeded that fold’s own target. The pooled error minus the target was not resolved from zero for any method. S-risk at the relative target. Folds over their own target count certified outer folds whose realised error exceeded that fold’s r*. | Method | Folds certified | Accepted | Error among accepted | Target r*, weighted over folds | Certified folds over their own target | Error minus target | | --- | --- | --- | --- | --- | --- | --- | | BETA · 7 outer folds · 70 people | | | | | | | | Standard CCA | 7 / 7 | 6,751 / 11,200 (coverage 60.3%) | 17.0% (95% interval 13.3%–21.6%) | 18.5% | 4 / 7 | −1.5 pp (95% interval −5.3 to +3.2 pp) no difference resolved | | CBraMod | 4 / 7 | 522 / 11,200 (coverage 4.7%) | 33.1% (95% interval 18.7%–48.5%) | 34.9% | 2 / 4 | −1.8 pp (95% interval −16.2 to +13.6 pp) no difference resolved | | EEGNet | 7 / 7 | 5,605 / 11,200 (coverage 50.0%) | 18.9% (95% interval 14.9%–23.7%) | 23.3% | 2 / 7 | −4.4 pp (95% interval −8.5 to +0.5 pp) no difference resolved | | EEGMAT · 5 outer folds · 36 people | | | | | | | | Spectral ridge | 0 / 5 | 0 / 2,160 (nothing accepted) | — (not defined: nothing accepted) | — (no fold certified) | — (no fold certified) | — (not defined: nothing accepted) | | EEGNet | 0 / 5 | 0 / 2,160 (nothing accepted) | — (not defined: nothing accepted) | — (no fold certified) | — (no fold certified) | — (not defined: nothing accepted) | | LaBraM · head only, encoder frozen | 0 / 5 | 0 / 2,160 (nothing accepted) | — (not defined: nothing accepted) | — (no fold certified) | — (no fold certified) | — (not defined: nothing accepted) | | LaBraM · LoRA rank 4 + head | 0 / 5 | 0 / 2,160 (nothing accepted) | — (not defined: nothing accepted) | — (no fold certified) | — (no fold certified) | — (not defined: nothing accepted) | Each fold holds about ten people, so a fold’s error also carries sampling noise; the count of folds over target is descriptive. ### A personal recalibration costs each new person labels R0 uses the population temperature and no label from the new person; R1 fits a personal temperature on a labelled prefix of that person’s own trials. With 40 labels per new person on BETA, NLL and AURC fell for every method. With 10 per person on EEGMAT, NLL rose, with the interval excluding zero, for three of the four methods: many personal temperatures were degenerate, and they were kept, as the protocol requires. Re-setting only the coverage cut-off from that prefix, without its labels, moved EEGMAT coverage further from the 80% target with the 10-label prefix, resolved for three of the four methods. R1 minus R0 on the same remaining trials; negative is better. | Method | Labelled prefix | Labels per new person | Change in NLL, R1 − R0 | Change in ECE | Change in AURC | People with a degenerate personal temperature | Label-free cut-off: change in distance from the 80% target | | --- | --- | --- | --- | --- | --- | --- | --- | | BETA | | | | | | | | | Standard CCA | block 1, one trial per target | 40 | −0.065 (95% interval −0.103 to −0.032) difference resolved | −0.019 (95% interval −0.048 to +0.009) no difference resolved | −0.012 (95% interval −0.022 to −0.005) difference resolved | 1 / 70 | +1.0 pp (95% interval −4.0 to +2.2 pp) no difference resolved | | CBraMod | block 1, one trial per target | 40 | −0.113 (95% interval −0.171 to −0.048) difference resolved | −0.083 (95% interval −0.120 to −0.042) difference resolved | −0.052 (95% interval −0.070 to −0.031) difference resolved | 0 / 70 | −0.3 pp (95% interval −4.4 to +3.8 pp) no difference resolved | | EEGNet | block 1, one trial per target | 40 | −0.080 (95% interval −0.135 to −0.020) difference resolved | −0.089 (95% interval −0.125 to −0.048) difference resolved | −0.016 (95% interval −0.027 to −0.004) difference resolved | 0 / 70 | +2.6 pp (95% interval −2.9 to +4.0 pp) no difference resolved | | EEGMAT | | | | | | | | | Spectral ridge | first 5 windows of each recording, then a one-window gap | 10 | +0.054 (95% interval +0.017 to +0.103) difference resolved | +0.059 (95% interval +0.006 to +0.084) difference resolved | +0.006 (95% interval −0.027 to +0.039) no difference resolved | 9 / 36 | +10.6 pp (95% interval +4.6 to +15.8 pp) difference resolved | | Spectral ridge | first 10 windows of each recording, then a one-window gap | 20 | +0.023 (95% interval +0.003 to +0.047) difference resolved | +0.031 (95% interval −0.006 to +0.061) no difference resolved | +0.013 (95% interval −0.022 to +0.048) no difference resolved | 9 / 36 | +1.7 pp (95% interval −2.0 to +5.7 pp) no difference resolved | | EEGNet | first 5 windows of each recording, then a one-window gap | 10 | +0.333 (95% interval −0.051 to +0.881) no difference resolved | −0.040 (95% interval −0.091 to +0.024) no difference resolved | −0.086 (95% interval −0.154 to −0.014) difference resolved | 8 / 36 | +5.8 pp (95% interval −1.1 to +10.3 pp) no difference resolved | | EEGNet | first 10 windows of each recording, then a one-window gap | 20 | +0.027 (95% interval −0.088 to +0.185) no difference resolved | −0.051 (95% interval −0.095 to +0.025) no difference resolved | −0.104 (95% interval −0.171 to −0.030) difference resolved | 5 / 36 | +1.7 pp (95% interval −3.2 to +6.0 pp) no difference resolved | | LaBraM · head only, encoder frozen | first 5 windows of each recording, then a one-window gap | 10 | +0.047 (95% interval +0.016 to +0.082) difference resolved | +0.057 (95% interval +0.012 to +0.082) difference resolved | −0.017 (95% interval −0.052 to +0.018) no difference resolved | 10 / 36 | +7.2 pp (95% interval +2.1 to +11.1 pp) difference resolved | | LaBraM · head only, encoder frozen | first 10 windows of each recording, then a one-window gap | 20 | −0.003 (95% interval −0.020 to +0.019) no difference resolved | +0.004 (95% interval −0.033 to +0.032) no difference resolved | −0.029 (95% interval −0.064 to +0.008) no difference resolved | 10 / 36 | +2.7 pp (95% interval −0.7 to +6.0 pp) no difference resolved | | LaBraM · LoRA rank 4 + head | first 5 windows of each recording, then a one-window gap | 10 | +0.937 (95% interval +0.184 to +1.924) difference resolved | +0.036 (95% interval −0.008 to +0.078) no difference resolved | −0.019 (95% interval −0.075 to +0.035) no difference resolved | 9 / 36 | +6.9 pp (95% interval +2.8 to +10.5 pp) difference resolved | | LaBraM · LoRA rank 4 + head | first 10 windows of each recording, then a one-window gap | 20 | −0.022 (95% interval −0.053 to +0.019) no difference resolved | −0.020 (95% interval −0.047 to +0.022) no difference resolved | −0.072 (95% interval −0.118 to −0.022) difference resolved | 4 / 36 | +1.8 pp (95% interval −2.7 to +5.4 pp) no difference resolved | ### LoRA and probability quality, on EEGMAT On EEGMAT, LaBraM adapted with rank-4 LoRA had a lower NLL than the head-only arm of the same recipe: calibrated −0.035 (95% interval −0.054 to −0.013). For ECE, no difference resolved (+0.015; −0.017 to +0.045). The LoRA arm is also more accurate, 64.4% against 56.9% balanced accuracy (seed 20260922, as the adaptation release publishes it; the three-seed means are on the model-adaptation page), so the NLL gap alone is not evidence of more honest probabilities. The adaptation itself is on [the model-adaptation page](https://bci.report/topics/model-adaptation/). ### Probability quality, with everything accepted Negative log-likelihood (NLL), expected calibration error (ECE, 15 equal-mass bins) and the Brier score after temperature scaling, and NLL and ECE before it where the model outputs a probability. Lower is better throughout. NLL and the Brier score depend on the number of classes, so EEGMAT (2) and BETA (40) are not compared. Binned ECE is biased upward under person resampling, so a small ECE can sit at the lower end of its interval. In one outer fold of the LaBraM head-only arm the population temperature reached its search bound; it was flagged and kept. Full coverage, 95% person-bootstrap intervals. | Method | NLL, calibrated | ECE, calibrated | Brier score, calibrated | NLL, raw | ECE, raw | | --- | --- | --- | --- | --- | --- | | EEGMAT | | | | | | | Spectral ridge | 0.679 (95% interval 0.668–0.691) | 0.037 (95% interval 0.033–0.070) | 0.486 (95% interval 0.475–0.498) | — (not applicable: no raw probability, not zero) | — (not applicable: no raw probability, not zero) | | EEGNet | 0.648 (95% interval 0.613–0.693) | 0.099 (95% interval 0.059–0.151) | 0.450 (95% interval 0.422–0.485) | 0.874 (95% interval 0.684–1.102) | 0.135 (95% interval 0.088–0.200) | | LaBraM · head only, encoder frozen | 0.683 (95% interval 0.674–0.691) | 0.036 (95% interval 0.034–0.072) | 0.490 (95% interval 0.481–0.498) | 0.683 (95% interval 0.662–0.703) | 0.061 (95% interval 0.046–0.090) | | LaBraM · LoRA rank 4 + head | 0.648 (95% interval 0.624–0.675) | 0.051 (95% interval 0.040–0.094) | 0.456 (95% interval 0.433–0.481) | 0.647 (95% interval 0.607–0.692) | 0.043 (95% interval 0.038–0.088) | | BETA | | | | | | | Standard CCA | 1.476 (95% interval 1.247–1.716) | 0.050 (95% interval 0.028–0.079) | 0.504 (95% interval 0.437–0.572) | — (not applicable: no raw probability, not zero) | — (not applicable: no raw probability, not zero) | | CBraMod | 2.809 (95% interval 2.638–3.001) | 0.119 (95% interval 0.084–0.154) | 0.843 (95% interval 0.815–0.871) | — (not applicable: no raw probability, not zero) | — (not applicable: no raw probability, not zero) | | EEGNet | 1.912 (95% interval 1.670–2.177) | 0.136 (95% interval 0.103–0.168) | 0.607 (95% interval 0.554–0.663) | 1.907 (95% interval 1.665–2.171) | 0.136 (95% interval 0.103–0.168) | Robustness and exploratory arms, each read on its own #### ds003810, motor imagery and rest: crude 10 people, two per test fold. Its contrasts below are secondary. The certified risk accepted nothing here for any method. | Method | Balanced accuracy, everything accepted | Accepted at confidence ≥ 0.8 | People with nothing accepted | L − S, AURC | L − S, error at 80% | Folds certified (S-risk) | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge | 54.9% (the published value) | 0 / 1,200 (coverage 0.0%) | 10 (of 10 people) | +0.003 (95% interval −0.031 to +0.032) no difference resolved | +0.5 pp (95% interval −1.1 to +2.3 pp) no difference resolved | 0 / 5 (nothing accepted) | | CBraMod | 62.9% (the published value) | 7 / 1,200 (coverage 0.6%) | 6 (of 10 people) | +0.013 (95% interval −0.014 to +0.030) no difference resolved | +0.7 pp (95% interval −0.8 to +2.4 pp) no difference resolved | 0 / 5 (nothing accepted) | | EEGNet | 71.0% (the published value) | 250 / 1,200 (coverage 20.8%) | 0 (of 10 people) | +0.038 (95% interval +0.009 to +0.061) difference resolved | +0.9 pp (95% interval −0.4 to +2.4 pp) no difference resolved | 0 / 5 (nothing accepted) | #### Secondary seeds Every other saved seed of the trained methods, each with its own calibration, cut-offs and selector, scored on that seed’s saved test predictions. | Method | Dataset | Seed | L − S, AURC | L − S, error at 80% | | --- | --- | --- | --- | --- | | EEGNet | EEGMAT | 20260920 | +0.007 (95% interval −0.029 to +0.041) no difference resolved | +1.4 pp (95% interval −0.6 to +3.2 pp) no difference resolved | | EEGNet | EEGMAT | 20260921 | −0.001 (95% interval −0.045 to +0.038) no difference resolved | +2.3 pp (95% interval 0.00 to +4.6 pp) no difference resolved | | LaBraM · head only, encoder frozen | EEGMAT | 20260923 | +0.044 (95% interval +0.013 to +0.074) difference resolved | +1.9 pp (95% interval +0.3 to +3.7 pp) difference resolved | | LaBraM · head only, encoder frozen | EEGMAT | 20260924 | +0.024 (95% interval −0.001 to +0.050) no difference resolved | +1.4 pp (95% interval +0.1 to +2.8 pp) difference resolved | | LaBraM · LoRA rank 4 + head | EEGMAT | 20260923 | +0.021 (95% interval −0.001 to +0.043) no difference resolved | +2.0 pp (95% interval +0.7 to +3.2 pp) difference resolved | | LaBraM · LoRA rank 4 + head | EEGMAT | 20260924 | +0.024 (95% interval −0.004 to +0.050) no difference resolved | +1.7 pp (95% interval +0.6 to +3.2 pp) difference resolved | | EEGNet | BETA | 20260919 | −0.002 (95% interval −0.011 to +0.007) no difference resolved | +0.9 pp (95% interval −0.1 to +1.8 pp) no difference resolved | | EEGNet | BETA | 20260920 | −0.006 (95% interval −0.013 to +0.002) no difference resolved | +0.3 pp (95% interval −0.5 to +1.0 pp) no difference resolved | | EEGNet | BETA | 20260921 | −0.010 (95% interval −0.019 to −0.001) difference resolved | +0.1 pp (95% interval −0.8 to +0.8 pp) no difference resolved | #### Two sensitivity arms: certified folds only The matched-model holdout trains the model on three quarters of the training people and calibrates it on the remaining quarter, so calibration and test scores come from one model. Platt scaling on two classes and vector scaling on BETA replace temperature scaling and can change predicted labels. | Method | Dataset | Arm | Folds certified | Certified folds over their own target | | --- | --- | --- | --- | --- | | Spectral ridge | EEGMAT | matched-model holdout | 0 / 5 | — (no fold certified) | | EEGNet | EEGMAT | matched-model holdout | 1 / 5 | 0 / 1 | | LaBraM · head only, encoder frozen | EEGMAT | matched-model holdout | 0 / 5 | — (no fold certified) | | LaBraM · LoRA rank 4 + head | EEGMAT | matched-model holdout | 0 / 5 | — (no fold certified) | | Standard CCA | BETA | matched-model holdout | 7 / 7 | 3 / 7 | | CBraMod | BETA | matched-model holdout | 6 / 7 | 3 / 6 | | EEGNet | BETA | matched-model holdout | 7 / 7 | 1 / 7 | | Spectral ridge | EEGMAT | Platt scaling | 0 / 5 | — (no fold certified) | | EEGNet | EEGMAT | Platt scaling | 0 / 5 | — (no fold certified) | | LaBraM · head only, encoder frozen | EEGMAT | Platt scaling | 0 / 5 | — (no fold certified) | | LaBraM · LoRA rank 4 + head | EEGMAT | Platt scaling | 0 / 5 | — (no fold certified) | | Standard CCA | BETA | vector scaling | 7 / 7 | 5 / 7 | | CBraMod | BETA | vector scaling | 4 / 7 | 2 / 4 | | EEGNet | BETA | vector scaling | 7 / 7 | 2 / 7 | #### A reject head trained with its classifier (exploratory) A SelectiveNet-style EEGNet, trained with its own selection head g (alpha 0.5, lambda 32, target coverage 0.8). Within the same model, g is set against the calibrated confidence of its own classifier head; a positive difference means g is worse. | Dataset | g − confidence, AURC | g − confidence, error at 80% | | --- | --- | --- | | EEGMAT | +0.011 (95% interval −0.012 to +0.035) no difference resolved | +2.4 pp (95% interval +0.8 to +4.2 pp) difference resolved | | BETA | +0.080 (95% interval +0.061 to +0.099) difference resolved | +1.9 pp (95% interval +1.0 to +2.6 pp) difference resolved | A cross-day label-budget arm on BNCI2015-001 stays unpublished while its source is held (see [the holds register](https://bci.report/releases/#hold-bnci2015-001-crossday)). ### Limits of route 1 ### Method comparisons Every protocol here is balanced or uniform by design: EEGMAT is half rest and half arithmetic, BETA uniform over its 40 targets, ds003810 balanced. The figures compare methods; none is a deployment error rate. On EEGMAT each class is its own recording, so recording-level differences are part of what any model or selector can use. BETA is cue-paced spelling with no idle state, so it gives no false-activation figure of any kind. ### Nothing accepted, no error rate A risk–coverage curve exists only where something is accepted. At zero coverage there is no error rate, not a zero one, and fewer than 10 accepted trials are flagged unstable. ### A nominal guarantee A certified selective risk assumes that calibration and test data are exchangeable. Across people they are not, so the guarantee is nominal: how often it fails is measured here, not assumed. ### Out-of-distribution input Rejecting likely errors on known classes is not detecting unfamiliar input: an uncertainty threshold may not flag an unseen class. Route 1 makes no out-of-distribution claim. [Mulder et al. (2026), motor imagery: arXiv:2603.13324 ↗](https://arxiv.org/abs/2603.13324) ### Calibration on fewer people Calibration scores come from inner models that see about three quarters of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. ### Published accuracies Temperature scaling keeps each predicted label, so every full-coverage accuracy equals the site’s published value. The matched-model, map-sensitivity and jointly trained arms use other models or maps; their accuracies are not published and are never mixed with the site’s. ### The difference rule A contrast is called a difference only when its paired 95% person-bootstrap interval excludes zero, with no multiplicity correction; otherwise “no difference resolved”. One row is above another only where their 95% intervals do not overlap; descriptive levels carry no verdict. ### Idle left out On the idle protocol, route 1 gives per-window proportions at each decision time, never a false-activation rate per hour, in a different unit from this page’s idle figures, so they are left out. ### ds003810 is crude Ten people, two per test fold: bootstrap intervals are crude, and no person-level spread is published. ### The recordings and the methods ### EEGMAT Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. [Dataset record ↗](https://physionet.org/content/eegmat/1.0.0/) · [Open Data Commons Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/) ### BETA Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. [Dataset record ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ### ds003810 Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ [Dataset record ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) #### Methods these results rest on, checked against their records on 4 October 2026 - [Guo, Pleiss, Sun & Weinberger, ICML 2017 · On calibration of modern neural networks ↗](https://arxiv.org/abs/1706.04599) - [Geifman & El-Yaniv, 2017 · Selective classification for deep neural networks ↗](https://arxiv.org/abs/1705.08500) - [Geifman & El-Yaniv, ICML 2019 · SelectiveNet: a deep neural network with an integrated reject option ↗](https://proceedings.mlr.press/v97/geifman19a.html) - [Geifman, Uziel & El-Yaniv, ICLR 2019 · Bias-reduced uncertainty estimation for deep neural classifiers (AURC) ↗](https://arxiv.org/abs/1805.08206) - [Nixon et al., 2019 · Measuring calibration in deep learning ↗](https://arxiv.org/abs/1904.01685) - [Mulder, Valdenegro-Toro, Sburlea & de Jong, 2026 · The challenge of out-of-distribution detection in motor imagery BCIs ↗](https://arxiv.org/abs/2603.13324) Checks behind these figures: both stage-0 audits passed. The independent primary audit passed 11 of its 12 checks; the one it failed found a documentation error, not a computation error — the run log said a degenerate personal temperature falls back to the population one, while the code keeps it, as the protocol requires, and every number was computed and reproduced that way. The independent secondary audit passed all 15. Both re-implement the definitions without the study code; they stay in the private run root and are pinned by hash. Data source: [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) · schema bci-report-reliable-decisions-update-v1. A longer write-up of the BETA results is on Hugging Face: [When should an EEG decoder abstain? ↗](https://huggingface.co/blog/Twu31/when-should-an-eeg-decoder-abstain) Research plan · three routes run ## Jev-style decision models for EEG A decoder has more choices than a class label: it can act, wait for more evidence, or ask for calibration. Status: all three routes have been run. The first route’s results are in the section above; the second’s and the third’s are on their own pages. “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. The first route needs no new model: it rescores saved outputs of models this site already publishes. The second trained models that share one encoder (small EEGNets from scratch, and CBraMod frozen or adapted by LoRA) and are given each question by an identifier, never in language. The third asks the questions in words, through a frozen text encoder and a small head over frozen EEG features. - First · Run · results above Reliable decisions Compare a fixed threshold, a simple statistical policy and a learned reject option on the same data. Report error at matched coverage, acceptance per class, and probability quality — Brier score, log loss and reliability — with calibration data kept apart from test data. Recalibration is a separate experiment, with its label cost counted. Published in the section above: error at matched coverage, log loss, the Brier score and calibration error. Not published: acceptance per class and reliability-diagram bins. [Reliable decisions: the results ↑](https://bci.report/topics/when-not-to-act/#reliable-decisions) - Then · Run · results on their own page One representation, several questions On the motor-imagery and sleep protocols separately, compare independent task models, a shared encoder with fixed heads, and the same encoder with a question-conditioned head, at matched data and compute. Run on motor imagery (OpenBMI) and sleep (BOAS, with EESM19 as a crude replication). One secondary arm, CBraMod adapted by LoRA on sleep, ran after the other results were known; its design was fixed beforehand. [One model, several questions: the results →](https://bci.report/topics/shared-encoder/) - Later · Run · results on their own page Questions in language Compare task identifiers, label templates and natural-language descriptions. Paraphrases and unseen concepts are tested separately and never merged into one zero-shot number. Run on SSVEP (BETA) and sleep (BOAS), with EESM19, OpenBMI motor imagery and Wearable-102 as secondary data. The pre-run permutation check failed twice and the owner revised it each time, the second time overriding its own “no further revision”; both failures are disclosed with the results. Questions about negation, location or waveform shape need signal-level ground truth. [Questions in language: the results →](https://bci.report/topics/questions-in-language/) All three routes, with one answer each: [Jev-style questions on EEG: the evidence →](https://bci.report/jev-style/) ### Limits of the existing data The sleep and P300 protocols on this site use balanced subsets. A probability estimated on them reflects that balance, so they serve to compare methods; a deployment error rate needs data in which each class occurs as often as in real use. The first input with explicit non-control states is the asynchronous SSVEP data set above: a small pilot kept as the development set, then the rest of the same release, scored under rules fixed in advance. ### Published work External evidence as its authors report it, checked against the original records on 27 September 2026. #### [Geifman & El-Yaniv, 2017 · Selective classification for deep neural networks ↗](https://arxiv.org/abs/1705.08500) **Shows** A trained classifier can be given a reject option with a target risk set by the user, trading coverage for error. **Limits** Evidence from image benchmarks. Its guarantee assumes test data that resemble the calibration data, which EEG across people and sessions often does not. #### [Geifman & El-Yaniv, ICML 2019 · SelectiveNet ↗](https://proceedings.mlr.press/v97/geifman19a.html) **Shows** Classification and rejection can be trained together, with risk at a chosen coverage as the explicit objective. **Limits** Not EEG evidence. Its guarantees do not carry over to a new participant or device by default. #### [Chatterjee et al., 2026 · Making conformal predictors robust in healthcare settings ↗](https://arxiv.org/abs/2602.19483) **Shows** On EEG seizure classification, patient shift left standard conformal prediction short of its nominal coverage, and personalised calibration raised coverage substantially. **Limits** One clinical task. It shows that coverage must be measured under the split that will actually be used. #### [Dai et al., 2026 · MTEEG: unified multi-task EEG analysis with low-rank adaptation ↗](https://arxiv.org/abs/2604.25131) **Shows** One pretrained EEG model can be adapted to several tasks at once with task-specific LoRA modules. **Limits** A separate adapter per task is not the same as several questions sharing one computation over the same window. #### [Jiang et al., 2024 · NeuroLM: a universal multi-task foundation model for EEG ↗](https://arxiv.org/abs/2409.00101) **Shows** EEG tokens fed to a language model, with instruction tuning, support several EEG tasks in one model. **Limits** Instruction-style EEG models already exist. A multi-task result is not evidence that arbitrary questions generalise. #### [Lu et al., 2025 · UniMind: LLMs for unified multi-task brain decoding ↗](https://arxiv.org/abs/2506.18962) **Shows** A language-model-based EEG model can select task-aware queries to decode several tasks. **Limits** Its queries are internal tokens, not criteria a user writes. #### [Gijsen & Ritter, 2024 · EEG-language pretraining for label-efficient clinical phenotyping ↗](https://arxiv.org/abs/2409.07480) **Shows** EEG aligned with clinical reports enables zero-shot classification and retrieval in the authors’ clinical evaluations. **Limits** Hospital EEG and clinical phenotypes. Nothing yet about wearable devices, few channels or arbitrary concepts. #### [Yu & Yao, 2026 · Visual Jev: Accurate and Efficient Decisions from Shared Visual Context ↗](https://arxiv.org/abs/2609.25845) **Shows** In vision, one encoding of an image and its context can serve several independent forced-choice questions run as a batch: the same encode-once, answer-many idea, applied to images. This plan started from it. **Limits** Its own typed-head control gave no consistent accuracy advantage, and its training gains were concentrated on the task families represented in training. Its speed-ups were measured on images. ### Where this connects [How much calibration?](https://bci.report/topics/calibration-budget/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *How often does an EEG decoder fire when nobody is giving a command?* https://bci.report/topics/when-not-to-act/ Figures from releases [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`extension-update-20261002`](https://bci.report/releases/#extension-update-20261002) (2026-10-02), [`context-update-20260927`](https://bci.report/releases/#context-update-20260927) (2026-09-27), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/when-not-to-act/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Why can a sleep stager be right most of the time and still miss whole stages? Sleep staging · two simple baselines Five-stage sleep staging of 30-second epochs against the publisher’s stored consensus hypnogram, in two Dreem cohorts, each analysed as its own experiment. - [Reviewed aggregate JSON ↓](https://bci.report/data/large-source-update.json) - Measured on: [Dreem Open Datasets (DOD-H and DOD-O)](https://bci.report/datasets/dreem-dod/) ## Short answer Because accuracy counts epochs, and sleep stages are far from equal in size: a stager can miss a whole stage and lose little accuracy. In two Dreem cohorts — **25** healthy sleepers and **55** people with obstructive sleep apnoea — a simple spectral ridge scored **72.3%** accuracy but **56.7%** balanced accuracy in the first and **72.5%** accuracy but **49.3%** balanced accuracy in the second, and never predicted N1 in either. A prior that always answers N2, the most frequent stage, sets the floor: it already reaches **48.0%** and **49.2%** accuracy, yet only **20.2%** and **20.5%** balanced accuracy. Experiment 1 of 2 · DOD-H ## DOD-H: healthy sleepers Recorded at French Armed Forces Biomedical Research Institute (IRBA), Fatigue and Vigilance Unit, Brétigny-sur-Orge, France. 25 nights, one per person, with 24,662 eligible 30-second epochs. 5 folds within the cohort: each tests 5 nights with models fitted on the other 20. **Mean over 25 nights, each counting once.** Accuracy and balanced accuracy for both baselines. Dot: the mean; line: whole-night bootstrap 95% interval. - Training prior · Accuracy: 48.0% (44.6%–51.4%) - Training prior · Balanced accuracy: 20.2% (20.0%–20.6%) - Spectral ridge · Accuracy: 72.3% (67.8%–76.3%) - Spectral ridge · Balanced accuracy: 56.7% (53.1%–60.0%) Both baselines, all four metrics. | Baseline | Accuracy | Balanced accuracy | Macro F1 | Cohen’s kappa | | --- | --- | --- | --- | --- | | Training prior — predicts N2 for every epoch · the floor | 48.0% (95% interval 44.6%–51.4%) | 20.2% (95% interval 20.0%–20.6%) | 13.0% (95% interval 12.3%–13.8%) | 0.000 (95% interval 0.000–0.000) | | Spectral ridge — standardised one-hot ridge, alpha 1, on 25 relative spectral values | 72.3% (95% interval 67.8%–76.3%) | 56.7% (95% interval 53.1%–60.0%) | 53.3% (95% interval 49.0%–57.3%) | 0.584 (95% interval 0.526–0.636) | **+36.5 pp** Spectral ridge minus training prior, balanced accuracy, the same 25 nights ### Well above the floor the prior sets Paired interval +32.9 to +39.8 pp: it excludes zero. Spectral ridge, stage by stage: mean over the nights on which each value is defined. Epochs: that stage’s count in the consensus. | Stage | Recall | Precision | F1 | | --- | --- | --- | --- | | Wake — 3,037 epochs | 64.8% (95% interval 54.9%–73.9%) | 76.4% (95% interval 70.1%–82.3%) | 66.5% (95% interval 58.5%–73.8%) | | N1 — 1,505 epochs | 0.0% (95% interval 0.0%–0.0%) (never predicted, so every N1 epoch is missed) | — (not defined: N1 was never predicted on any of the 25 nights) | 0.0% (95% interval 0.0%–0.0%) | | N2 — 11,879 epochs | 85.4% (95% interval 77.4%–91.8%) | 77.3% (95% interval 72.8%–81.8%) | 78.9% (95% interval 74.0%–83.1%) | | N3 — 3,514 epochs | 59.4% (95% interval 46.8%–71.2%) (on 24 of 25 nights; 1 without N3 in the consensus) | 69.7% (95% interval 55.7%–81.9%) (on 24 of 25 nights; 1 where N3 was never predicted) | 56.2% (95% interval 43.6%–67.7%) | | REM — 4,727 epochs | 73.6% (95% interval 63.8%–82.6%) | 63.8% (95% interval 55.4%–71.5%) | 64.6% (95% interval 56.1%–72.2%) | The ridge predicted N1 on none of the 25 nights: its N1 recall and F1 are a measured zero, and its N1 precision is undefined, because precision divides by the number of epochs predicted as N1, which is zero. Experiment 2 of 2 · DOD-O ## DOD-O: people with obstructive sleep apnoea Recorded at Stanford Sleep Medicine Center, United States (clinical trial NCT03657329). 55 nights, one per person, with 53,161 eligible 30-second epochs. 5 folds within the cohort: each tests 11 nights with models fitted on the other 44. **Mean over 55 nights, each counting once.** Accuracy and balanced accuracy for both baselines. Dot: the mean; line: whole-night bootstrap 95% interval. - Training prior · Accuracy: 49.2% (46.3%–52.2%) - Training prior · Balanced accuracy: 20.5% (20.1%–21.2%) - Spectral ridge · Accuracy: 72.5% (69.7%–75.0%) - Spectral ridge · Balanced accuracy: 49.3% (47.3%–51.2%) Both baselines, all four metrics. | Baseline | Accuracy | Balanced accuracy | Macro F1 | Cohen’s kappa | | --- | --- | --- | --- | --- | | Training prior — predicts N2 for every epoch · the floor | 49.2% (95% interval 46.3%–52.2%) | 20.5% (95% interval 20.1%–21.2%) | 13.3% (95% interval 12.8%–13.9%) | 0.000 (95% interval 0.000–0.000) | | Spectral ridge — standardised one-hot ridge, alpha 1, on 25 relative spectral values | 72.5% (95% interval 69.7%–75.0%) | 49.3% (95% interval 47.3%–51.2%) | 47.4% (95% interval 44.9%–49.7%) | 0.542 (95% interval 0.504–0.578) | **+28.8 pp** Spectral ridge minus training prior, balanced accuracy, the same 55 nights ### Well above the floor the prior sets Paired interval +26.7 to +30.7 pp: it excludes zero. Spectral ridge, stage by stage: mean over the nights on which each value is defined. Epochs: that stage’s count in the consensus. | Stage | Recall | Precision | F1 | | --- | --- | --- | --- | | Wake — 10,427 epochs | 80.4% (95% interval 75.5%–85.1%) | 76.5% (95% interval 72.0%–80.7%) | 75.6% (95% interval 71.7%–79.3%) | | N1 — 2,860 epochs | 0.0% (95% interval 0.0%–0.0%) (never predicted, so every N1 epoch is missed) | — (not defined: N1 was never predicted on any of the 55 nights) | 0.0% (95% interval 0.0%–0.0%) | | N2 — 26,271 epochs | 92.8% (95% interval 90.5%–94.7%) | 71.0% (95% interval 67.9%–74.0%) | 79.8% (95% interval 77.3%–82.0%) | | N3 — 5,500 epochs | 22.1% (95% interval 15.4%–29.2%) (on 52 of 55 nights; 3 without N3 in the consensus) | 63.4% (95% interval 52.4%–73.9%) (on 49 of 55 nights; 6 where N3 was never predicted) | 27.9% (95% interval 20.3%–35.8%) (on 53 of 55 nights; 2 with N3 neither in the consensus nor predicted) | | REM — 8,103 epochs | 49.5% (95% interval 43.1%–55.9%) (on 53 of 55 nights; 2 without REM in the consensus) | 68.2% (95% interval 60.9%–74.9%) (on 54 of 55 nights; 1 where REM was never predicted) | 52.5% (95% interval 46.0%–58.7%) (on 54 of 55 nights; 1 with REM neither in the consensus nor predicted) | The ridge predicted N1 on none of the 55 nights: its N1 recall and F1 are a measured zero, and its N1 precision is undefined, because precision divides by the number of epochs predicted as N1, which is zero. Cohorts and design ## What was evaluated, and how Nights evaluated 80 81 archive records, minus 1 the publisher declares has no consensus scoring: DOD-H 25, DOD-O 55. Eligible 30-second epochs 77,823 77,901 in all, minus 3 unscored and 75 with a zero or invalid channel scale. Models fitted 20 Two cohorts × 5 folds × two baselines; none failed. ### Target The consensus hypnogram as the publisher stores it, in five stages — Wake, N1, N2, N3, REM — for 30-second epochs. The individual scorers’ votes were not reconstructed. ### Signal and features Five EEG derivations (C3-M2, F3-F4, F3-M2, F3-O1, F4-O2) at 250 Hz. Each epoch and channel is divided by its own largest absolute value; Welch spectra then give delta, theta, alpha, sigma and beta power as a log share of 0.5–30 Hz power: 25 values per epoch. ### Two baselines The training prior uses the training folds’ stage frequencies, the same for every epoch, so it predicts N2 everywhere. The spectral ridge is a standardised one-hot ridge regression with alpha 1 on the 25 values. Nothing was tuned, calibrated on held-out data, oversampled or class-balanced, and no run was retried on its outcome; stages keep their natural prevalence. ### Folds Five folds within each cohort, assigned by a hash of each record before any outcome was seen. Every night is tested once per baseline, by a model fitted on the other four folds of its own cohort. ### Metrics Computed per night, then averaged so each night counts once, however long. Balanced accuracy is the mean recall over the stages a night contains; macro F1 the mean F1 over the stages it contains or the model predicts. Cohen’s kappa is zero by construction for a prior that gives every epoch the same answer. ### Uncertainty Pointwise 95% whole-night bootstrap within each cohort: 10,000 draws (PCG64, seed 20261003), the same draws for both baselines and every metric; a paired difference is published for balanced accuracy only. Conditional on the fixed cross-validation predictions — refitting and fold assignment are not included — and not adjusted for multiple comparisons. Held · no score ## Neural-network and foundation models: held The stored signal metadata says millivolts, while the publisher’s own converter treats the same arrays as microvolts. Foundation models and other amplitude-sensitive encoders are held because the source’s physical units disagree. The two baselines above divide each epoch and channel by a positive gain before computing relative spectra, so they do not depend on that unit. [Its row in the holds register →](https://bci.report/releases/#hold-dreem-amplitude-sensitive-models) Methods & limits ## Limits ### Two separate experiments DOD-H and DOD-O were recorded at different centres with different equipment. Their figures are not a controlled comparison of health status, and nothing here measures transfer from one cohort to the other. ### Not clinical, not a diagnosis These are research baselines; nothing here is evidence of equivalence with human experts. ### Physical units unresolved The relative spectra do not depend on the unit, but that does not establish that calibration, reference, clipping or filtering are equivalent. ### Uncalibrated probabilities The ridge’s scores are not calibrated probabilities, and no probability score is published. ### A floor, not a chance level Balanced accuracy averages only the stages a night contains. The prior predicts N2 everywhere, so on a night that lacks one of the five stages its balanced accuracy is one quarter rather than one fifth; that is why it sits slightly above one fifth. ### Specific to this setup Do not compare these figures with papers that use other channels, cohorts, preprocessing or splits. ### Dreem Open Datasets (DOD-H and DOD-O) Antoine Guillot, Fabien Sauvet, Emmanuel H. During and Valentin Thorey · Dreem Open Datasets: Multi-Scored Sleep Datasets to Compare Human and Automated Sleep Staging, IEEE Transactions on Neural Systems and Rehabilitation Engineering (2020), doi:10.1109/TNSRE.2020.3011181; preprint arXiv:1911.03221. Data: Dreem Open Datasets, Zenodo, doi:10.5281/zenodo.15900394. Official repository: github.com/Dreem-Organization/dreem-learning-open, revision 8b332d6827f5ae6a22f4bb97b4deef4273238ec3. Deposit licence: MIT, as declared on the Zenodo record. BCI Report redistributes no upstream code or data. [Paper ↗](https://doi.org/10.1109/TNSRE.2020.3011181) · [Preprint ↗](https://arxiv.org/abs/1911.03221) · [Zenodo deposit ↗](https://zenodo.org/records/15900394) · [Official repository at the pinned revision ↗](https://github.com/Dreem-Organization/dreem-learning-open/tree/8b332d6827f5ae6a22f4bb97b4deef4273238ec3) · [MIT](https://opensource.org/licenses/MIT) ### Consent and ethics, cohort by cohort **DOD-H** · Ethics approval: Approved by the Committees of Protection of Persons (CPP); declared to the French National Agency for Medicines and Health Products Safety; carried out in compliance with the French Data Protection Act, International Conference on Harmonization (ICH) standards and the Declaration of Helsinki (1964, as revised in 2013). Informed consent: not stated. The paper's DOD-H description contains no informed-consent sentence. **DOD-O** · Ethics approval: not stated. No ethics committee or institutional review board is named for DOD-O in the paper. Informed consent: All trial participants gave informed written consent before taking part. Source of these statements: arXiv:1911.03221v4 (27 April 2020), II. Materials and Methods, A. Datasets, read 2026-10-03. Data source: [large-source-update.json](https://bci.report/data/large-source-update.json) · schema bci-report-large-source-update-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Why can a sleep stager be right most of the time and still miss whole stages?* https://bci.report/topics/sleep-staging/ Figures from release [`large-source-update-20261003`](https://bci.report/releases/#large-source-update-20261003) (2026-10-03). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/sleep-staging/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Can resting-state EEG separate Parkinson's disease from controls? Clinical research · resting-state EEG Parkinson's disease against controls, person by person: a spectral model on resting-state EEG, and a comparator given only age and sex; and an infant case-control pilot with age and recording-setting controls; and resting EEG in schizophrenia and controls beside age and sex. - [Reviewed aggregate JSON ↓](https://bci.report/data/clinical-update.json) - Measured on: [OpenNeuro ds004584 · Parkinson's disease, rest eyes open](https://bci.report/datasets/ds004584/) ## Short answer Above chance, as a group-level research result from one site. Across **149** people with Parkinson's disease or in the control group, a relative-spectrum logistic regression on resting-state EEG separated the two groups at **71.7%** balanced accuracy (95% interval **63.5%–79.3%**, AUROC **0.76**, chance **50%**); a comparator given only age and sex reached **55.0%** (**46.9%–63.6%**) on the same folds. **The same 149 people and the same 5 folds.** Person-level balanced accuracy, Parkinson's disease against controls. Dot: the estimate; line: descriptive bootstrap 95% interval; dashed line: chance, 50%. - Relative-spectrum logistic regression — resting-state EEG: 71.7% (63.5%–79.3%) - Age-and-sex comparator — no EEG input: 55.0% (46.9%–63.6%) **+16.6 pp** EEG minus the age-and-sex comparator, same 149 people ### A higher mean, with overlapping intervals The EEG model's interval starts at 63.5%; the comparator's ends at 63.6%. Research results on a public data set. Not a diagnosis, not diagnostic accuracy, and not a medical device or medical advice. EEG model and comparator ## EEG against age and sex Two classes, chance 50%; 5 stratified participant folds. Person-level balanced accuracy with its 95% interval, macro F1 and AUROC. | Row | Inputs | Balanced accuracy | Macro F1 | AUROC | | --- | --- | --- | --- | --- | | Relative-spectrum logistic regression — log relative band power, 60 shared channels | EEG only | 71.7% (63.5%–79.3%) | 71.1% | 0.76 | | Age and sex only — not an EEG model | no EEG | 55.0% (46.9%–63.6%) | 54.0% | 0.56 | Cohort and features ## One site, 149 people, two minutes each The 60 named channels shared by four source layouts; the first 120 seconds of each recording, cut into 30 disjoint four-second windows; log relative power in five bands spanning 1–45 Hz. Each person's 300 features are averaged within that person, so the classifier sees one vector per person rather than 4,470 independent examples. Scaling is fitted on training folds only, the classifier is a fixed balanced logistic regression (C=1), and there is one split seed with no held-out tuning. People 149 100 people with Parkinson's disease and 49 controls, one site Participant folds 5 each person held out once Four-second windows 4,470 averaged within each person Methods & limits ## Limits ### Not a diagnosis Group-level discrimination on a fixed research split, reported as balanced accuracy. Screening, clinical validation and interpretation for any individual person are outside its scope. ### Not replicated One cohort, one split seed. The intervals are descriptive bootstrap ranges, because the cross-validation training folds overlap. ### Age and sex are not removed The comparator shows that part of this separation is available with no brain signal at all. The EEG model is not adjusted for age or sex, so its score is not a measure of disease-specific brain activity. ### Group differences are not only in the brain Anything that differs between patients and controls travels with the labels: medication, recording session, movement, alertness. A case/control score cannot separate those from the disease. ### Foundation models: held The source states no physical amplitude unit. A relative spectrum needs none; models that need a calibrated scale are held. ### Independently replayed 149 recordings and 4470 spectral windows with a separate FFT, all classifiers refit. Reconstructed features and metrics differed by 0; intervals differed at floating-point precision. ### OpenNeuro ds004584 · Rest eyes open, Parkinson's disease and controls Dataset: Arun Singh, Rachel Cole, Arturo Espinoza, Jim Cavanagh and Nandakumar Narayanan · Rest eyes open, OpenNeuro ds004584 v1.0.0, doi:10.18112/openneuro.ds004584.v1.0.0. Cohort study: Arun Singh, Rachel C. Cole, Arturo I. Espinoza, Jan R. Wessel, James F. Cavanagh and Nandakumar S. Narayanan · Evoked mid-frontal activity predicts cognitive dysfunction in Parkinson's disease, Journal of Neurology, Neurosurgery & Psychiatry 94, 945-953 (2023), doi:10.1136/jnnp-2022-330154. Transport mirror: NEMAR on004584, doi:10.82901/nemar.on004584. Every participant gave written informed consent, with decisional capacity established, in accordance with the Declaration of Helsinki; all research protocols were approved by the University of Iowa Human Subjects Review Board (IRB 201707828). Only cohort aggregates appear here: no participant rows, no clinical or cognitive scores, and no per-group age or sex composition. [OpenNeuro dataset ↗](https://doi.org/10.18112/openneuro.ds004584.v1.0.0) · [Cohort study ↗](https://doi.org/10.1136/jnnp-2022-330154) · [Open manuscript ↗](https://europepmc.org/article/PMC/PMC10592174) · [NEMAR mirror ↗](https://doi.org/10.82901/nemar.on004584) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) Data source: [reviewed aggregate JSON](https://bci.report/data/clinical-update.json) · schema bci-report-clinical-update-v1. Infant case-control pilot · awake EEG ## Do EEG case-control results survive age and recording-setting controls? IESSEEG is a public EEG dataset from one institution’s retrospective cohort: 100 people, 50 cases of infantile epileptic spasms syndrome, each on the recording made before the treatment the source study evaluated, and 50 controls. Each person contributes one clinician-selected awake segment, chosen by a fixed rule with no signal-driven choice, read for its first 60 s on the 17 channels every recording shares. Every arm is scored out of fold on the publisher’s 5 participant-disjoint folds: 80 people train, 20 are held out, and each person is held out once. **The probes and every score were reproduced independently; the NeuroLM features themselves were not re-extracted by a second implementation.** Spectral EEG reached 81.00% balanced accuracy, and 80.00% with age added. Age alone reached 68.00%, recording settings alone 61.00%, and the training prior 50.00%. With age regressed out inside each training fold, spectral EEG reached 67.00%. On the same people, folds and probe, frozen GPT hidden states reached 84.00% and the frozen tokenizer 77.00%; with age regressed out, 71.00% and 72.00%. **Both tables.** Point estimates out of fold; no interval or test was specified. One awake recording per person, first 60 s. Age and recording settings differ between groups; regressing out age is a sensitivity analysis. Balanced accuracy is the primary metric; AUROC is shown beside it for every arm, and no arm or threshold is chosen from these results. The age ≤2 column re-scores the same out-of-fold predictions for the people aged 2 or under, with nothing refitted; it is a descriptive subgroup. Only the band’s size and its case and control counts are published; per-group age means and sex counts stay unpublished. In this cohort the band holds 68 people: 42 cases and 26 controls. ### Age and acquisition controls for an awake EEG case-control pilot The spectral arms read Welch spectra at each recording’s native rate (4 s windows overlapping by half, 0.25 Hz resolution): log relative power in five bands from 1 to 45 Hz and log total power over the same range, for each of the 17 channels, 102 features per person. Every learned arm uses the same probe, fixed in advance with no tuning: standardisation fitted on the training fold, then L2 logistic regression (C = 1). The age terms are log1p of age at the recording and its square; the residualized arm regresses each feature on them inside the training fold and applies those coefficients to the training and held-out people. The recording-settings arm reads the original scalp channel count, 19 in 59 recordings and 17 in 41, and the native sampling rate, 200 Hz in 98 and 2,000 Hz in 2. Six arms with and without EEG, each scored out of fold | Arm | Balanced accuracy, all 100 | AUROC, all 100 | Balanced accuracy, age ≤2 years — 68 people | | --- | --- | --- | --- | | Without EEG | | | | | Training prior — No input: the training fold’s share of cases, which calls every person a case | 50.00% | 0.500 | 50.00% | | Age only — No EEG: log1p of age at the recording, and its square | 68.00% | 0.621 | 50.00% | | Acquisition metadata only — No EEG: the original scalp channel count and sampling rate | 61.00% | 0.524 | 69.23% | | With EEG | | | | | EEG spectral features — The 102 log band-power features of the 17 channels | 81.00% | 0.824 | 76.56% | | EEG spectral features + age — The 102 spectral features and the two age terms | 80.00% | 0.843 | 73.90% | | EEG after training-fold age residualization — The 102 spectral features with age regressed out inside each training fold | 67.00% | 0.703 | 69.41% | ### Do frozen EEG representations survive age controls? NeuroLM-B is read frozen, as a feature extractor: its weights are unchanged, and no text is involved. Each recording’s 60 s are cut into 15 windows of 4 s, with one token per channel and second, 68 tokens per window. Two of the model’s internal states are read: the tokenizer’s, after its final normalisation, and the GPT branch’s hidden states, after the final layer normalisation, 768 values each, averaged over each window’s tokens and then over the windows. The probe and the training-fold age regression are those of the spectral arms, unchanged. Four arms reading frozen NeuroLM-B states with the same probe, on the same people and folds | Arm | Balanced accuracy, all 100 | AUROC, all 100 | Balanced accuracy, age ≤2 years — 68 people | | --- | --- | --- | --- | | Frozen EEG tokenizer — The 768 frozen tokenizer values of each person | 77.00% | 0.901 | 73.90% | | Tokenizer + training-fold age residualization — The tokenizer values with age regressed out inside each training fold | 72.00% | 0.798 | 71.34% | | Frozen GPT hidden states — The 768 frozen GPT hidden-state values of each person | 84.00% | 0.902 | 84.43% | | GPT hidden states + training-fold age residualization — The GPT hidden-state values with age regressed out inside each training fold | 71.00% | 0.794 | 75.92% | **NeuroLM.** NeuroLM’s preprocessing differs from the spectral pipeline: the two 2,000 Hz recordings are resampled to 200 Hz, then a 0.1–75 Hz band-pass, 50 and 60 Hz notches and division by 100. A difference between the two pipelines therefore mixes preprocessing with architecture. The checkpoint’s training stage and its prior exposure to these participants are unknown. [NeuroLM source code at revision `0cda9876` ↗](https://github.com/935963004/NeuroLM/tree/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177) · checkpoint SHA-256 `ffe098bc138b89f8817d3710a3604498d8ecd15135080e2ca27735d05c6d29ab` Research results on a public data set. Not a diagnosis, not diagnostic accuracy, and not a medical device or medical advice. Limits of this pilot ### One small development cohort A small awake-only case-control development cohort from one institution. Labels are each participant’s case or control status, with no timed spasm events. There is no external or clinical validation, and the publisher’s full benchmark is outside this pilot. ### Regressing out age Regressing out age inside each training fold assigns no share of the score to age, does not show that confounding is gone, and can also remove disease-related signal that varies with age. ### Recording settings Recording settings (channel count and sampling rate) are a no-EEG diagnostic control, shown beside the EEG rows and kept out of model comparisons. ### NeuroLM, frozen NeuroLM is used as a frozen feature extractor read by supervised logistic probes: its weights are unchanged, and there is no text prompt, text generation or zero-shot diagnosis. ### Independently reproduced Spectral and control arms: 30 models (25 refits and 5 prior checks) and 600 held-out probabilities. An independent program refit every learned model, checked every prior, rebuilt every held-out probability with zero difference, and recomputed the features of one recording per channel-count and sampling-rate group with its own EDF reader and FFT. NeuroLM arms: 20 probe refits, 400 held-out probabilities, 300 input files and 100 feature files. An independent program refit every probe and rebuilt every held-out probability with zero difference, checked every input file’s hash, and checked each feature file’s hash, layout and finiteness. An independent read-only audit checked every input file’s hash, timing and calibration: 300 files, 2,884,494,142 bytes; all 100 recordings passed, with none held and none with gaps. Endpoint and constant-sample counts were kept as descriptions; no recording was excluded on them. ### IESSEEG · awake case-control pilot Dataset: IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome, Hugging Face Capur/IESSEEG at revision 1c32ee3a, CC BY 4.0 for the dataset and annotations. The release card declares that the 400 case clinician-selected segments were previously analysed by Rajaraman et al. (2024), Computational EEG attributes predict response to therapy for epileptic spasms. The release card states institutional review board approval of the retrospective study, with waivers of informed consent and HIPAA research authorization, and asks that no one attempt to re-identify participants. The publication boundary is the one the Parkinson’s comparison uses: aggregate scores only, with no participant rows, predictions, ages or per-group age and sex composition; the age ≤2 column gives the band’s size and its case and control counts. The card gives the approval as protocol 17-000026 and does not name the institution whose board approved it. The participant-level crosswalk between that analysis and this release is unresolved; the cohort adds no dataset to any count. [Dataset at the pinned revision ↗](https://huggingface.co/datasets/Capur/IESSEEG/tree/1c32ee3a425a324808a9d275b60ce85625af35c8) · [Release card ↗](https://huggingface.co/datasets/Capur/IESSEEG/blob/1c32ee3a425a324808a9d275b60ce85625af35c8/README.md) · [LICENSE ↗](https://huggingface.co/datasets/Capur/IESSEEG/blob/1c32ee3a425a324808a9d275b60ce85625af35c8/LICENSE) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Data use](https://bci.report/data-use/#sources-clinical-iesseeg) Data source: [reviewed aggregate JSON](https://bci.report/data/clinical-iesseeg-update.json) · schema bci-report-clinical-iesseeg-update-v1. Schizophrenia case-control comparison · resting EEG ## Does resting EEG separate schizophrenia from controls better than age and sex do? ASEEG is a public resting-EEG dataset from one clinic: 101 people with eyes-closed recordings, 51 with schizophrenia and 50 controls, and eyes-open recordings for 97 of the same people (47 and 50). Each recording is read for its first 256 s on its 19 channels as released. Every arm is scored out of fold on 5 stratified participant folds, averaged over 10 split seeds. In the eyes-closed recordings, EEG spectra reached 71.19% balanced accuracy and the comparator given only age and sex 71.24%: a difference of −0.05 pp (−11.51 to +11.86 pp). EEG spectra with age and sex reached 75.94%, a difference of +4.71 pp from that comparator (−4.40 to +13.86 pp). With age and sex regressed out inside each training fold, EEG spectra reached 69.81% (62.27%–77.11%). Seed-averaged out-of-fold scores with descriptive 95% paired participant-bootstrap intervals; chance is 50.00% for balanced accuracy and 0.5 for AUROC. Balanced accuracy is the primary metric; AUROC is shown beside it for every arm. ### Eyes closed, 101 people Four arms on the eyes-closed recordings | Arm | Balanced accuracy (95% interval) | AUROC | | --- | --- | --- | | Without EEG | | | | Only age and sex — No EEG: age in years and sex | 71.24% (95% interval 62.61%–79.71%) | 0.694 | | With EEG | | | | EEG spectra — The 95 log relative band-power features | 71.19% (95% interval 63.29%–78.71%) | 0.775 | | EEG spectra + age and sex — The 95 spectral features, age and sex | 75.94% (95% interval 68.64%–82.81%) | 0.833 | | EEG spectra after age-and-sex residualisation — The 95 spectral features with age and sex regressed out inside each training fold | 69.81% (95% interval 62.27%–77.11%) | 0.775 | Paired contrasts | Contrast | Role | Difference in balanced accuracy (95% interval) | Interval excludes zero | | --- | --- | --- | --- | | EEG spectra − Only age and sex — The interval includes zero; the difference is not an equivalence test. | Primary | −0.05 pp (95% interval −11.51 to +11.86 pp) | No | | EEG spectra + age and sex − Only age and sex — The interval includes zero. | Secondary | +4.71 pp (95% interval −4.40 to +13.86 pp) | No | | EEG spectra after age-and-sex residualisation − chance — The interval lies above zero; regressing out linear effects inside training folds does not show that confounding is gone. | Sensitivity | +19.81 pp (95% interval +12.27 to +27.11 pp) | Yes | ### Frozen CBraMod, eyes closed One arm reading frozen CBraMod with the same probe | Arm | Balanced accuracy (95% interval) | AUROC | | --- | --- | --- | | Frozen CBraMod — The person-mean of CBraMod’s frozen embeddings, read by the same probe | 76.65% (95% interval 69.50%–83.37%) | 0.826 | Paired contrasts | Contrast | Role | Difference in balanced accuracy (95% interval) | Interval excludes zero | | --- | --- | --- | --- | | Frozen CBraMod − Only age and sex — The interval includes zero; the encoder has no age or sex input. | Frozen encoder | +5.41 pp (95% interval −4.92 to +15.93 pp) | No | | Frozen CBraMod − EEG spectra — Descriptive; the interval includes zero. | Descriptive | +5.46 pp (95% interval −1.40 to +12.36 pp) | No | CBraMod is read as a frozen encoder by the same person-level probe. Its contract differs from the foundation-model panels: the curators’ 128 Hz recordings are resampled to 200 Hz; each channel’s mean is removed in every 2-second window; the channels are in the released, undocumented order; the probe reads the mean of 128 window embeddings per person; and the microvolt scale comes from the descriptor’s Figure 2 caption, checked by a label-free plausibility test set in advance; the files store no unit. The checkpoint predates the dataset’s release and its stated pretraining corpus is TUEG; a difference between CBraMod and the spectral probe mixes preprocessing with architecture. ### Eyes open, 97 of the same people Two arms on the eyes-open recordings | Arm | Balanced accuracy (95% interval) | AUROC | | --- | --- | --- | | Only age and sex, eyes open — No EEG: age in years and sex | 69.91% (95% interval 60.96%–78.68%) | 0.676 | | EEG spectra, eyes open — The 95 features of the eyes-open recording | 70.67% (95% interval 63.08%–77.93%) | 0.723 | Paired contrasts | Contrast | Role | Difference in balanced accuracy (95% interval) | Interval excludes zero | | --- | --- | --- | --- | | EEG spectra, eyes open − Only age and sex, eyes open — The interval includes zero. | Second condition | +0.75 pp (95% interval −10.64 to +11.90 pp) | No | These are the same people as above, so this is not an independent replication. Research results on a public data set. Not a diagnosis, not diagnostic accuracy, and not a medical device or medical advice. Limits of this section One clinic and one device; the recordings were taken retrospectively from the clinic’s records, and the controls were people having routine EEG there for reasons unrelated to mental health. A group-level research comparison. The data descriptor reports that the groups differ in sex; the age-and-sex comparators and the residualised arm are shown, and residualising removes linear effects inside training folds only. Medication and clinical scores are not released, so the patients’ treatment travels with the label. The curators band-passed the signals to 1–45 Hz, notched them, downsampled them to 128 Hz and removed recordings with severe artefacts or incomplete sessions; transient artefacts remain by design. The released demographic sheet differs slightly from the descriptor’s summary and is used as released; ages and sexes pair with recordings by the shared subject number, which the arms combining EEG with age and sex rely on. The channel order is undocumented, so no channel-specific reading is possible. Every value is the mean over ten split seeds of five-fold participant-held-out cross-validation; intervals are descriptive paired participant-bootstrap ranges that ignore the overlap of training folds and carry no multiplicity correction. ### How this differs from the Parkinson’s and IESSEEG sections - ten split seeds averaged - 10,000 bootstrap draws - the first 256 s in 64 four-second windows - 128 Hz as released, with 2-second Welch segments and 0.5 Hz bins - 19 channels on the curators’ ear reference, in an undocumented order - five relative bands, no log total power - age in years and sex as the confounds, with no recording-setting arm - each channel’s mean removed in every 2-second CBraMod window, and no age-and-sex-residualised CBraMod arm ### Independently reproduced - **Pass**: The arms, contrasts, folds, split seeds and bootstrap were fixed in a protocol that an independent check passed before any score using the real labels existed. - **Pass**: The out-of-fold predictions were sealed by SHA-256 before the independent audit read them. - **Pass**: An independent program re-derived every spectral feature from the raw files, refitted every arm and re-extracted every CBraMod embedding, with zero difference. - **Pass**: An independent reader checked every reading printed here against the frozen reading rules. ### ASEEG · Atieh Schizophrenia EEG Bagherzadeh S, Norouzi M, Farokhshad F, Tolou Koroushi P, Ghasri A, Bahri Hampa S, Kazemi R, Rostami R, Shalbaf A. ASEEG: Atieh Schizophrenia EEG. Zenodo, doi:10.5281/zenodo.18029536 (First version, 2025-12-23), CC BY 4.0. Data descriptor: Scientific Data 13:1211 (2026), doi:10.1038/s41597-026-07968-9. Modifications: the first 256 s of each released recording read as released; log relative band power and frozen CBraMod features; cohort-level scores only. Derived analysis by BCI Report; not endorsed by the authors. The data descriptor states that the ethics review board at Shahid Beheshti University of Medical Sciences approved the study, under ethics code IR.SBMU.MSP.REC.1404.441, and that specific informed consent covered the recording and the sharing of de-identified data, with written authorization from a parent, guardian or responsible adult where needed. The publication boundary is the one the Parkinson’s and IESSEEG sections use: aggregate scores only, with no participant rows or predictions and no per-group age or sex composition. **The data descriptor says:** “The participant group was sourced from the Atieh Clinic, a private facility located in Tehran, Iran.” “EEG signals were obtained retrospectively from the clinic’s medical records.” “All studies and activities are conducted under the ethics code number IR.SBMU.MSP.REC.1404.441 and adhere to the principles of ethical research.” “specific informed consent was obtained for the EEG recording procedure and for the subsequent sharing of de-identified data.” “a parent, legal guardian, or other directly responsible adult provided written authorization on a dedicated consent form.” The data descriptor is CC BY-NC-ND 4.0, so only short quotes appear here. CBraMod: code MIT, weights Apache-2.0 as its official model card states; not redistributed; Wang et al., arXiv 2412.07236. [Zenodo record ↗](https://doi.org/10.5281/zenodo.18029536) · [Data descriptor ↗](https://doi.org/10.1038/s41597-026-07968-9) · [CBraMod, arXiv ↗](https://arxiv.org/abs/2412.07236) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Data use](https://bci.report/data-use/#sources-clinical-aseeg) Data source: [reviewed aggregate JSON](https://bci.report/data/clinical-aseeg-update.json) · schema bci-report-clinical-aseeg-update-v1. Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Can resting-state EEG separate Parkinson's disease from controls?* https://bci.report/topics/clinical-groups/ Figures from releases [`clinical-aseeg-update-20261106`](https://bci.report/releases/#clinical-aseeg-update-20261106) (2026-11-06), [`clinical-iesseeg-update-20261020`](https://bci.report/releases/#clinical-iesseeg-update-20261020) (2026-10-20), [`clinical-update-20260923`](https://bci.report/releases/#clinical-update-20260923) (2026-09-23). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/clinical-groups/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # What do the files show once a public EEG dataset is checked end to end? Data quality and reproducibility · seven public EEG datasets, read from their files Shin2022, Kim beta-range and eldBETA are worked cases: their files counted and checked first, then a fixed baseline beside its reference. MIND is a conditional case on record IDs, with two controls that read no EEG. Damm and NeBULA each have one participant run and publish counts and file facts; ZuCo 2.0 carries a timing hold. - [Reviewed aggregate JSON ↓](https://bci.report/data/data-quality-update.json) - Measured on: [Shin2022 · closed-loop motor imagery EEG (Figshare 20383716)](https://bci.report/datasets/shin2022/); [Kim beta-range SSVEP · 40 targets, 14.0–21.8 Hz (Figshare 28806815)](https://bci.report/datasets/kim2025-beta-range/) ## Short answer Shin2022’s files repeat some runs byte for byte, and some feedback windows end early or carry a flat channel; on the windows that remain, a participant-held-out check put alpha and beta band power at **52.59%** balanced accuracy, beside **52.33%** for a control that reads only how long each trial lasted and **50.00%** for the training prior. Training-free CCA reached **53.40%** on Kim beta-range against chance of **2.50%**. Across all **100** eldBETA participants every file passes the checks, and training-free CCA reached **74.38%** mean participant accuracy and filter-bank CCA **87.10%**, against chance of **11.11%**. In one converted ZuCo 2.0 run some fixation times fall outside the recording, so no EEG–text score is computed. Worked case · Shin2022 · closed-loop cursor control, dry electrodes ## What survives a participant-held-out check? Right-versus-left instructed target in the first second of cursor feedback, for a participant held out entirely: each of the 10 participants is left out once, and a ridge fitted on the other 9 scores their windows. On the signal-valid windows counted below, alpha and beta band power reached 52.59% participant-mean balanced accuracy (51.51%–53.66%). A control that reads only each trial’s full feedback duration reached 52.33% (50.87%–54.02%). The training prior gives 50.00%. Participant-mean balanced accuracy: the 10 held-out participants’ balanced accuracies averaged with equal weight, with its 95% participant-bootstrap interval. The pooled value counts every signal-valid window together: a secondary figure, without an interval. | Arm | Participant-mean balanced accuracy | Pooled (secondary) | What it reads | | --- | --- | --- | --- | | Participant-weighted training prior — reference | 50.00% (95% interval 50.00%–50.00%) | 50.00% | No input: always the larger training class, with each training participant weighted equally. | | Normalized spectral ridge — EEG | 52.59% (95% interval 51.51%–53.66%) | 52.55% | 32 log band powers relative to each channel’s 1–40 Hz power: alpha (8–12 Hz) and beta (13–30 Hz) on each of the 16 stored channels; ridge with alpha 1.0, no tuning. | | Retrospective control | | | | | Retrospective duration control | 52.33% (95% interval 50.87%–54.02%) | 52.15% | One value: the log of the full feedback duration; ridge with alpha 1.0. | The scores cover the signal-valid windows only: feedback lasting at least one second, and a first second that is finite with positive 1–40 Hz power on all 16 channels. The label is the instructed target in closed-loop cursor recordings, which carry visual, controller and behaviour confounds. The duration control reads the whole feedback duration, known only after the trial ends, so it checks a confound; the spectral arm sits close to it, and no test between arms was run. Ten participants, one recording day each; the intervals are descriptive. ### From the runs in the release to the scored windows Each count is read from a hash-pinned record (the label and duplicate audit, the signal census review or the scored export), and each step follows from the one above it. Right and left are the instructed targets. | Step | Count | Right · left | What the files show | | --- | --- | --- | --- | | Runs in the release | 110 | — | Runs in the release’s human recordings; the simulated branch is excluded and adds no participant. | | Byte-identical pairs | 9 | — | Pairs of runs that are byte-identical: each pair sits within one participant and joins a BW30 run and an NT0 run. | | Distinct runs | 101 | — | Distinct runs once each byte-identical pair counts once. They hold 2,412 closed feedback windows. | | Incomplete run — held | 1 | — | One run ends inside a feedback window; it is held whole. | | Complete runs | 100 | 12 · 12 per run | Complete distinct runs, each with twelve closed feedback windows per target. They cover all 10 participants. | | Closed feedback windows | 2,400 | 1,200 · 1,200 | Closed feedback windows in the complete runs. | | Shorter than one second — excluded | 167 | 76 · 91 | Feedback ended within the first second, so the window holds less than the one second the arms read. | | At least one second | 2,233 | 1,124 · 1,109 | Windows with at least one second of feedback. | | A flat channel in the first second — excluded | 13 | 7 · 6 | Windows whose first second is constant on at least one channel, which leaves that channel no 1–40 Hz power; the rule was frozen after a score-free census and before any fit. Flat on all channels: 8; on some: 5. Windows with non-finite input: 0. | | Signal-valid | 2,220 | 1,117 · 1,103 | Signal-valid windows: the population all three arms are scored on. That is 92.5% of the closed windows and 99.42% of those lasting at least one second; all 10 participants keep windows of both targets. | MOABB’s dataset summary gives a trial length of 6 s: that is the feedback cap (MaxFeedbackDuration) set in the header of each of the 110 runs read. MOABB’s trial and run counts for Shin2022 are findings on [its row in the open register](https://bci.report/datasets/register/#r-site-shin2022). ### How it was checked - **Held**: The first, strict protocol over all duration-eligible windows met a flat signal interval at its zero-power gate and stopped before any model was fitted; that attempt stays failed and held. - **Pass**: A census read the signal of every duration-eligible window, with no model and no score. - **Pass**: A second protocol with the signal-validity rule was frozen after the census and before any fit. - **Pass**: Predictions were sealed before the audit; metrics were computed only after it. - **Pass**: An independent audit rebuilt every metric from the sealed predictions. - **Pass**: An independent release review passed the aggregate values and the coverage for publication. ### Held or unestablished - The paper pairs NT0 with BW60, while the byte-identical runs pair BW30 with NT0; which recording NT0 stands for is unresolved. - Comparisons between the BW, CV and NT control-parameter conditions are held: one BW60 header records a 5-second buffer, and the NT and CV headers record 30-second defaults. - The incomplete run is held whole; using its closed windows would need its own frozen rule. - A task over every closed window, whatever its duration, would need its own frozen protocol. - The full scalp montage, the display latency, the absolute hardware calibration and any overlap with pretraining corpora are unestablished. - The physical cause of the flat intervals is unknown. - Exact deduplication leaves trial independence unestablished; every run of a participant stays in that participant’s fold. ### Source, licence and consent ### Shin2022 · Data from: Closed-loop motor imagery EEG simulation for brain-computer interfaces (Shin, Suma and He) Closed-loop right/left cursor control with visual feedback; the label is the instructed target. Figshare record 20383716, version 1, the only version the record lists. Hyonyoung Shin, Daniel Suma and Bin He · Data from: Closed-loop motor imagery EEG simulation for brain-computer interfaces, figshare, doi:10.6084/m9.figshare.20383716.v1, CC BY 4.0. Paper: Hyonyoung Shin, Daniel Suma and Bin He, Closed-loop motor imagery EEG simulation for brain-computer interfaces, Frontiers in Human Neuroscience 16:951591 (2022), doi:10.3389/fnhum.2022.951591. Funding: NIH AT009263. Modifications: canonical deduplication of the byte-identical runs, first-second window selection and signal-validity filtering. Derived analysis by BCI Report; not endorsed by the authors. The simulator code is a separate work linked from the data record; the data licence covers the data record. [Simulator code ↗](https://github.com/mcvain/bci-simulator) [Dataset record ↗](https://doi.org/10.6084/m9.figshare.20383716.v1) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Dataset page →](https://bci.report/datasets/shin2022/) ### Ethics and consent, as the source states them **Ethics approval:** The study was reviewed and approved by the Institutional Review Board of Carnegie Mellon University; the paper gives no approval number. **Informed consent:** The participants gave written informed consent to take part in the study. The data record also states: “All identifying information have been removed.” “Please cite the above paper if you use any data included in this dataset.” [Paper, with its ethics statement ↗](https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2022.951591/full) Worked case · Kim beta-range SSVEP · training-free CCA ## What does training-free CCA give on 40 frequencies in the beta range? Which of 40 flickering targets from 14.0 to 21.8 Hz a participant attends, from a 2-second window, with no training data. Standard CCA ran on every trial of all 40 participants (240 blocks, 9,600 trials) and reached 53.40% mean participant accuracy (45.89%–61.01%). Chance is 2.50%, one in 40 balanced targets. ### Intake: what the files hold All 40 task files passed their MD5 checks; the rest and questionnaire files were not acquired. The files store the targets in the order 14 + j + 0.2·i Hz, which differs from the ascending list in MOABB’s metadata, and the references follow the stored order. ### How it was checked - **Pass**: A separate review admitted the as-stored, scale-invariant CCA arm before any score. - **Pass**: Every file passed identity, shape, timing, frequency, phase, finite-value and rank checks before any score. - **Pass**: All 9,600 outputs were sealed before a separate process scored them; this is process separation, not blinded labels, because the stored class order is known. - **Pass**: An independent audit verified the counts and the bootstrap, and a separate CCA implementation reproduced every fixed fixture. - **Pass**: An independent review checked the figures and the wording against the audit aggregate. ### Caveats - The scores use amplitudes as stored. CCA does not depend on a common amplitude scale, but common channel gains, freedom from clipping and the saved reference history are not established, so models whose output depends on amplitude are held. - The onset is nominal: it is derived from the stored start of each epoch, 2 seconds before the cue, at 1,024 Hz, and no trigger trace is stored. - Each participant’s six blocks were recorded on one day. - The filter runs forward and backward inside each window, so the pipeline is offline and gives no response-time figure. - The pipeline is BCI Report’s fixed recipe; it does not reproduce the paper’s processing. - Frequencies, people and channels differ from BETA and Wearable-102, so comparing absolute scores across those datasets does not isolate the frequency range. - The accuracy interval is the participant-bootstrap interval of the difference from the fixed 2.5% chance level, shifted up by 2.5 points; no algorithm was run as chance and no test was run. - The independent audit reproduced 60 of 60 fixed fixtures; it did not rerun every trial. **Kim beta-range SSVEP · amplitude-dependent models · Held.** [Its entry in the holds register →](https://bci.report/releases/#hold-kim-amplitude-dependent-models) ### Source, licence and consent ### Kim beta-range SSVEP: 40 targets from 14.0 to 21.8 Hz (Kim, Won, Ahn and Jun, Scientific Data) Forty-target SSVEP speller selection with stimuli from 14.0 to 21.8 Hz, six blocks per person on one day. Figshare record 28806815, version 3, the latest of the 3 versions the record lists. Heegyu Kim · SSVEP 40class dataset, figshare, doi:10.6084/m9.figshare.28806815.v3, CC BY 4.0. Paper: Heegyu Kim, Kyungho Won, Minkyu Ahn and Sung Chan Jun, A 40-Class SSVEP Speller Dataset: Beta Range Stimulation for Low-Fatigue BCI Applications, Scientific Data 12:1751 (2025), doi:10.1038/s41597-025-06032-2. Modifications: the 40 task files only; the mean of M1 and M2 subtracted, 11 channels, a 2-second window and a band-pass filter inside it; no data redistributed. Derived analysis by BCI Report; not endorsed by the authors. The paper’s Data availability links version 2; the research used version 3 (the latest), which adds a questionnaire. Both versions carry CC BY 4.0 in the licence field. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.28806815.v3) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Dataset page →](https://bci.report/datasets/kim2025-beta-range/) ### Ethics and consent, as the source states them **Ethics approval:** “The Institutional Review Board at Gwangju Institute of Science and Technology (GIST) approved this experiment (20211201-HR-64-02-04)” **Informed consent:** “all subjects were informed of all experimental procedures and signed consent forms” The data descriptor also states: “All research data has been consented to by participants, with explicit agreement for public sharing.” The confidentiality clause of the consent forms excludes questionnaire-related personal data; the research acquired the 40 task files only and excluded the questionnaire and rest files. [Data descriptor, with its ethics statement ↗](https://www.nature.com/articles/s41597-025-06032-2) Conditional case · MIND · stored codes on record IDs ## What do stored markers support in a motor-imagery release? Stored class codes 4 versus 5 and 6 versus 7, each pair classified on its own, from the EEG of a four-class motor-imagery release; the codes are as stored and are not verified imagery. In the primary panel of 28 record IDs (84 runs, 3,360 trials), two whole runs train the pipeline and the third is held out, for each record ID and pair. CSP + LDA stands beside two controls that read no EEG: an ordinal lookup and the majority code. Record-ID mean balanced accuracy: out-of-fold trials pooled within each record ID, then record IDs weighted equally. | Stored codes | CSP + LDA | Ordinal lookup (no EEG) | Majority (no EEG) | | --- | --- | --- | --- | | 4/5 | 56.96% | 47.62% | 50.00% | | 6/7 | 51.85% | 45.24% | 50.00% | Each stored-code pair follows one of only two complementary 20-trial label schedules. In the primary panel, 6 record IDs for 4/5 and 5 for 6/7 repeat one schedule across all three runs; the others mix the two. Under the frozen ordinal lookup and its tie rule, this composition explains the ordinal values below 50%: they describe this lookup and schedule rather than a chance floor. No replacement control was fitted after the scores were seen. ### Four contrasts, with record-ID resampling ranges Descriptive sensitivity ranges for this fixed set of record IDs: the 2.5th and 97.5th percentiles of 20,000 paired record-ID resamples drawn from one shared index table. They are no confidence intervals and no test. | Stored codes | Contrast | Mean difference | Record-ID resampling range | | --- | --- | --- | --- | | 4/5 | CSP + LDA − ordinal lookup | +9.35 pp | −2.68 to +20.00 pp | | 4/5 | CSP + LDA − majority | +6.96 pp | +4.05 to +10.00 pp | | 6/7 | CSP + LDA − ordinal lookup | +6.61 pp | −4.05 to +15.95 pp | | 6/7 | CSP + LDA − majority | +1.85 pp | −0.48 to +4.23 pp | ### Secondary panel, apart The 2 record IDs whose first run was excluded (4 runs, 160 trials): one run trains the pipeline and the other is held out. Reported apart with no interval and never pooled with the primary panel; its schedule composition is not printed because with two record IDs it would show each one’s outcome. | Stored codes | CSP + LDA | Ordinal lookup (no EEG) | Majority (no EEG) | | --- | --- | --- | --- | | 4/5 | 53.75% | 50.00% | 50.00% | | 6/7 | 50.00% | 50.00% | 50.00% | ### From the runs in the release to the two panels | Step | Runs | Record IDs | Trials | What the files show | | --- | --- | --- | --- | --- | | Runs in the release | 90 | 30 | — | Runs in the release: three per record ID. | | Overfull first runs — excluded | 2 | — | — | First runs holding more class markers than the 40 a run requires, excluded whole before any waveform was read; no trial was salvaged from them. | | Runs evaluated | 88 | — | 3,520 | Runs that passed the frozen finite and non-degenerate checks, which is every one; the checks do not clear artifacts. | | Primary panel | 84 | 28 | 3,360 | Record IDs with three runs each: the primary panel. | | Secondary panel | 4 | 2 | 160 | The two record IDs whose first run was excluded keep two runs each: the secondary panel. | ### How it was checked - **Pass**: All out-of-fold outputs were sealed together before a separate scoring process. - **Pass**: An independent replay reproduced every hard output, decision value, metric and resampling output exactly. - **Pass**: An independent review approved the figures, the four contrasts, the schedule explanation and the wording. ### Caveats - Windows run from 1 to 9 seconds after each stored class marker. Most class markers have no cue marker before them and no cue marker sits exactly 2 seconds before its class marker, so the windows do not certify pure imagery. - The unit is the record ID. The release describes 30 participants, and whether record IDs and people match one to one was not reconciled, so no claim about people follows. - Pair 4/5 above the majority control does not exclude temporal, visual or movement explanations. - Each run is filtered whole, offline and in both directions, so no response-time figure follows. - The independent replay starts from the sealed epochs; raw preprocessing and the numerical checks were source-reviewed and bound by hash and were not recomputed. - The headers leave the reference field empty while the BIDS sidecars describe a parietal reference; the declared calibration and reference were kept as stored and were not verified physiologically. ### Source, licence and consent Licence statements, side by side | Source | Licence as stated | How it was read | | --- | --- | --- | | Paper | CC BY 4.0 | quoted from the research intake receipt of the paper page | | Dataset description in the archive | CC BY 4.0 | as the research protocol records the dataset description inside the downloaded archive | | MOABB adapter | CC BY-NC 4.0 | the licence field of the MOABB dataset adapter | | ScienceDB record page | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) | the schema.org metadata the record page serves | ScienceDB’s terms-of-use page was not read. Aggregate values only; no data are redistributed. ### MIND: A Multimodal fNIRS-EEG Dataset for Unilateral Limb Motor Imagery (ScienceDB 10.57760/sciencedb.34326) Four-class directional motor imagery of the right upper limb with EEG and fNIRS; here the EEG of stored codes 4 versus 5 and 6 versus 7. ScienceDB record 10.57760/sciencedb.34326, record version 4.0.0. Lufeng Feng, Baomin Xu, Haoran Zhang, Bihai Lin, Zuxuan Deng, Sidi Tao, Chenyu Liu, Shifan Jia, Li Duan and Ziyu Jia · MIND: A Multimodal fNIRS-EEG Dataset for Unilateral Limb Motor Imagery, Science Data Bank, doi:10.57760/sciencedb.34326. Paper: arXiv 2602.04299. Modifications: EEG only, filtered over each whole run, resampled and cut into stored-marker windows; no data redistributed. Derived analysis by BCI Report; not endorsed by the authors. [Dataset record ↗](https://doi.org/10.57760/sciencedb.34326) ### Ethics and consent, as the source states them **Ethics approval:** “The study protocol was approved by the Ethics Committee. All procedures followed institutional review board (IRB) guidelines and relevant privacy regulations.” **Informed consent:** “After confirming that they understood the instructions, all participants voluntarily signed written informed consent forms.” **Gaps:** Ethics and consent are quoted from the paper, which names no committee and gives no approval number. The intake receipt records no capture time. [Paper, with its ethics statement ↗](https://arxiv.org/html/2602.04299v1) Worked case · eldBETA · SSVEP selection by older adults · training-free CCA and filter-bank CCA ## What do training-free CCA and filter-bank CCA give across all of eldBETA? Which of nine flickering targets a participant attends, from a 3-second window, with no training data. Standard CCA and filter-bank CCA ran on every trial of all 100 published participants (700 blocks, 6,300 trials), with none excluded and none dropped. Mean participant accuracy was 74.38% for CCA (70.10%–78.63%) and 87.10% for filter-bank CCA (83.81%–90.14%), against chance of 11.11%; paired by participant, filter-bank CCA minus CCA was +12.71 pp (+10.41 to +15.05 pp). At least 12 of 63 correct is the one-sided 5% binomial threshold for one person: 98 of 100 participants reach it with CCA and 100 with filter-bank CCA. Filter-bank CCA is higher than CCA for 84 participants, equal for 11 and lower for 5. The 12-of-63 threshold is a per-person one-sided 5% binomial with no correction across the 100 people. The 11 participants counted as equal are at 63 of 63 under both decoders. ### Intake: what the files hold One GDF 2.11 recording per block, read with the header calibration applied; every archive was checked against its MD5, one of them against the value its depositor supplied; the derived MATLAB epochs were only checked for agreement and never analysed. The file checks cover every participant: 100 of 100 archives verified and readable, 700 blocks, and 6,300 of 6,300 trials pass on all 64 channels and on the 9 analysed channels, none dropped. The continuous recordings hold 21,675,520 non-finite values and none inside the analysed windows: one run in each of 665 of the 700 blocks, across all 64 channels, starting after the last analysed window and ending at the recording’s end inside its final one-second data record. This is consistent with padding of that last record, and the values are left as recorded. ### How it was checked - **Pass**: Every participant file was read and checked, with no score, before the protocol was frozen. - **Pass**: An independent review passed the frozen protocol before any score. - **Pass**: A bridge reproduced this page’s one-participant file counts and the earlier one-participant CCA count exactly. - **Pass**: Predictions were sealed without reading labels, then scored. - **Pass**: An independent audit re-cut all 6,300 windows from the raw archives, recomputed every prediction with its own code and reproduced the bootstrap exactly. - **Pass**: An independent review checked the figures and the wording against the audit aggregate. ### Caveats - Filter-bank CCA minus CCA compares two fixed recipes, seven sub-bands with five harmonics against unfiltered CCA with three, so it does not isolate the filter bank. - Filter-bank CCA ran with the code behind the filter-bank CCA method page, its filters designed at the recordings’ native 1,000 Hz; the lowest sub-band rings for about 3.5 seconds, longer than the 3-second window, so the value is not comparable with filter-bank CCA values on BETA. - The fifth harmonic of 10 Hz is 50 Hz, the mains frequency the amplifier notch-filters, and candidate harmonics coincide at 24 Hz for both decoders and at 36 and 40 Hz for filter-bank CCA; the references were left unchanged. - The window is 3 seconds from 140 samples after each event marker, an analysis convention rather than a measured latency; the filtering runs forward and backward inside each window, so the pipeline is offline and gives no response-time or information-transfer-rate figure. - Neither decoder depends on a per-channel amplitude scale. - The pipeline is BCI Report’s fixed recipe and does not reproduce the data descriptor’s processing. - Targets, people, window length, sampling rate and preprocessing differ from BETA and the other SSVEP cases, so absolute accuracies across datasets are not comparable. ### Held or unestablished - The derived MATLAB epochs were not used: their amplitude scale, resampling and exact endpoints are unestablished. - The raw recordings were read with the header calibration applied; the recorded unit code has no independent audit. eldBETA is in SingLEM’s published pretraining list, Table I of the SingLEM paper; BETA, a separate database, is absent from it, as SingLEM’s exposure row on BETA shows. [SingLEM paper, Table I ↗](https://arxiv.org/abs/2509.17920v3) · [SingLEM’s pretraining-exposure rows →](https://bci.report/methods/singlem/#pretraining-exposure) [CCA method page →](https://bci.report/methods/cca/) · [Filter-bank CCA method page →](https://bci.report/methods/fbcca/) ### Source, licence and consent ### eldBETA: A Large Eldercare-oriented Benchmark Database of SSVEP-BCI for the Aging Population (Liu, Wang, Gao and Chen) Nine-target SSVEP selection by older adults, recorded in seven blocks per person. Figshare record 18032669, version 3, the latest of the 3 versions the record lists. Bingchuan Liu, Yijun Wang, Xiaorong Gao and Xiaogang Chen · eldBETA: A Large Eldercare-oriented Benchmark Database of SSVEP-BCI for the Aging Population, figshare, doi:10.6084/m9.figshare.18032669.v3, CC BY 4.0. Data descriptor: Scientific Data, doi:10.1038/s41597-022-01372-9. Funding: National Natural Science Foundation of China (No. 62171473). Modifications: decoding of the original recording format with the header calibration; fixed channel and window selection; window-local filter-bank filtering for FBCCA; cohort-level accuracies published, no participant-level values. Derived analysis by BCI Report; not endorsed by the authors. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.18032669.v3) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ### Ethics and consent, as the source states them **Ethics approval:** The protocol followed the Declaration of Helsinki and was approved by the institutional review board of Tsinghua University, approval number 20210032. **Informed consent:** The participants were familiarised with the protocol and gave full written consent before the experiment. The data descriptor also states: “The eldBETA database is open-access for research and can be downloaded from the website” [Data descriptor, with its ethics statement ↗](https://www.nature.com/articles/s41597-022-01372-9) Status cases · one participant run in each ## Two cases with one participant run With one person every score is that person’s, so these two cases publish counts and file facts and hold every score. Status case · Damm finger motor imagery · random cue order ## One participant’s two random-order runs, end to end The runs hold 150 delivered cues over five fingers (thumb, index, middle, ring and little): 75 balanced trials in each held-out run. A fixed pipeline (log band power from an 8–30 Hz filter on each epoch, with fixed shrinkage LDA fitted on the training run only) was trained on one run and tested on the other, in both directions; eight variants and controls were fixed before scoring, and an independent audit rebuilt all 16 reports. ### What the files show Each random-order run is 916 seconds of 65 EDF channels at 512 Hz: 62 scalp channels, two reference channels and a marker channel. The event stream of each run holds 227 state blocks, and every cue follows a 3-second white screen and a 3-second fixation cross and lasts 6 seconds. The files carry no sample column, so cue samples are reconstructed from the stored onsets. ### Variants and controls, fixed before scoring - **Primary**: All 62 scalp channels, 1 to 5 seconds after the cue - **Window sensitivity**: All 62 scalp channels over the full cue interval, 0 to 6 seconds - **Precue control**: All 62 scalp channels in the 2 seconds before the cue - **Channel sensitivity**: A fixed central subset of 15 channels, 1 to 5 seconds after the cue - **Order control**: The trial index alone, with no signal input - **Order control**: The trial index modulo five, with no signal input - **Order control**: The previous delivered cue, with no signal input: privileged context that no live system would have - **Shuffled-label check**: The primary features with the training labels shuffled once with a fixed seed ### Held or unestablished - Cue samples are reconstructed from the stored onsets; physical trigger timing is not validated. - The EEG sidecars and the counterbalancing log disagree on which condition came first, so no order effect is inferred. - The sequential-order runs were not evaluated: their two annotation tables are byte-identical copies of one file, and no transfer between the random and sequential orders was tested. - The record lists 20 participants; the others are neither acquired nor evaluated. **Damm finger motor imagery · one-person score · Status only.** [Its entry in the holds register →](https://bci.report/releases/#hold-damm-one-participant-score) ### Source, licence and consent ### Damm finger motor imagery: Random and Sequential Order Finger Motor Imagery (OpenNeuro ds008446) Five-finger motor imagery cued in random and in sequential order; here the random-order runs only. OpenNeuro ds008446 at its latest snapshot, read at a pinned commit of its mirror. Laura Marie Damm, Dai Jiang and Andreas Demosthenous · Random and Sequential Order Finger Motor Imagery, OpenNeuro ds008446, doi:10.18112/openneuro.ds008446.v1.0.1, CC0. Funding: Engineering and Physical Sciences Research Council (EPSRC) grant EP/S022139/1. Modifications: the first participant’s two random-order runs only; no score from these data is published. Derived analysis by BCI Report; not endorsed by the authors. Licence: CC0, as dataset_description.json states it. The README cites a LICENSE file that the tree at the pinned commit does not hold; the only licence statement is CC0 in dataset_description.json. [Dataset record ↗](https://doi.org/10.18112/openneuro.ds008446.v1.0.1) ### Ethics and consent, as the source states them **Ethics approval:** “UCL Humanities, Arts and Sciences Research Ethics Committee; Approval Number: 26907/001” **Informed consent:** No consent statement in README.md, dataset_description.json, CHANGES or the OpenNeuro dataset metadata at the pinned commit; neither the README nor the metadata names a paper, so no paper was read. [Dataset description, with its ethics statement ↗](https://raw.githubusercontent.com/OpenNeuroDatasets/ds008446/8d8448e2dcd731b251f2f3c85b98cce5847cb000/dataset_description.json) Status case · NeBULA · reaching with exoskeleton assistance, EEG and EMG ## One participant’s reaches under three levels of assistance One participant reaching to three targets under free, low and high assistance: 89 delivered trials (29, 30 and 30), 127 EEG channels at 1,000 Hz. The EEG arms were trained on two assistance runs and tested on the third (15 fits), beside an order diagnostic with no signal input, a precue window and a label shuffle; EMG and EEG plus EMG arms (15 fits, no pretrained model) ran on the same trials. An independent audit rebuilt every arm. The paper and the Figshare record’s description state 40 participants; the record’s files hold 39 participant archives, and MOABB lists 39. The source describes executed reaching with an upper-limb exoskeleton; MOABB’s catalog lists the dataset under motor imagery. ### Arms run **EEG pilot:** - EEG, all 127 channels, 0 to 2 seconds after the Go cue - EEG, all 127 channels, the 500 milliseconds before the Go cue - EEG, a fixed set of central channels, 0 to 2 seconds after the Go cue - Order diagnostic: trial position and the previous delivered target, with no signal input; privileged context that no live system would have - The primary EEG features with the training labels shuffled once with a fixed seed **EMG and EEG plus EMG:** - EMG, 11 rows, 0 to 2 seconds after the Go cue - The 127 EEG features and the 11 EMG features concatenated in a fixed order - EMG, the 500 milliseconds before the Go cue - The EMG features with the training labels shuffled once - The concatenated features with the same shuffled training labels ### What the files show - The original markers hold 25 boundary ties where one trial’s end and the next trial’s start share a sample (7 free, 9 low, 9 high), with no backward overlap. The first parse stopped on them before any feature, model or score; a reviewed fix that accepts a shared boundary reran within the original budget. - The EMG files hold 11 channel rows where their metadata states 16; the rows were not relabelled or mapped to muscles. - All 356 stored markers of the 89 trials match the EEG marker files in label, order and sample position. - The EMG and EEG files of each run have equal sample lengths. ### Held or unestablished - Go is a cue and the reach marker marks reaching the target; neither certifies physical movement onset. - Stored marker agreement between the EEG and EMG files does not certify clock synchronisation or amplitude calibration. - The other participants’ files in the record are neither acquired nor evaluated. **NeBULA · one-person scores · Status only.** [Its entry in the holds register →](https://bci.report/releases/#hold-nebula-one-participant-scores) ### Source, licence and consent ### NeBULA (Neuromechanical Biomarkers for Upper Limb Assessment): an EEG-EMG dataset from a standardized reaching task (Garro et al., Scientific Data 12, 831) Reaching to three targets with an upper-limb exoskeleton under free, low and high assistance, with EEG and EMG. Figshare record 27301629, version 1, the only version the record lists. Florencia Garro, Elena Fenoglio, Indya Ceroni, Inna Forsiuk, Michele Canepa, Michael Mozzon, Agnese Bruschi, Francesco Zippo, Matteo Laffranchi, Lorenzo De Michieli, Stefano Buccelli, Michela Chiappalone and Marianna Semprini · An EEG-EMG Dataset from a Standardized Reaching Task for Biomarker Research in Upper Limb Assessment, figshare, doi:10.6084/m9.figshare.27301629.v1, CC BY 4.0. Paper: Florencia Garro, Elena Fenoglio, Indya Ceroni, Inna Forsiuk, Michele Canepa, Michael Mozzon, Agnese Bruschi, Francesco Zippo, Matteo Laffranchi, Lorenzo De Michieli, Stefano Buccelli, Michela Chiappalone and Marianna Semprini, An EEG-EMG dataset from a standardized reaching task for biomarker research in upper limb assessment, Scientific Data 12:831 (2025), doi:10.1038/s41597-025-05042-4. Modifications: the first participant’s files only; no score from these data is published. Derived analysis by BCI Report; not endorsed by the authors. The paper’s own licence, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, covers the data descriptor article only; the data licence is the Figshare record’s. The files’ own dataset description names Zenodo 13627307, which is no registered public record: the Zenodo API, the record page, DataCite and doi.org each return 404. The paper names Figshare alone. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.27301629.v1) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ### Ethics and consent, as the source states them **Ethics approval:** “The study’s experimental procedures were approved by the local ethical committee (CER Liguria:784/2022 - DB id 12044).” **Informed consent:** “All participants gave consent to participate and openly share the data collected.” [Data descriptor, with its ethics statement ↗](https://www.nature.com/articles/s41597-025-05042-4) Status case · ZuCo 2.0 on NEMAR · no score ## A timing hold on one converted run In one converted run of ZuCo 2.0 on NEMAR, some fixation times fall outside the recording. No EEG–text score is computed until the converter’s timing rules and a source-checksummed fixture are available. The hold covers that one converted run; no timing was corrected and no row removed. ### The hold clears with - the exact converter revision, configuration and environment, with its rules for casts, time origin and boundaries - a source-checksummed timing fixture with the original start and end timestamps and the recording’s start and end clocks - the canonical mapping of triggers, sentences and runs No consent or ethics statement for this dataset has been read yet. [Source and stated terms ↓](https://bci.report/topics/data-quality/#zuco2-source) **ZuCo on NEMAR · converted-run timing · No score.** [Its entry in the holds register →](https://bci.report/releases/#hold-zuco2-converted-run-timing) ## Each case in the open register Each case has a row in the open register, with the furthest step checked and the gaps that remain. - Shin2022: [Row in the open register →](https://bci.report/datasets/register/#r-site-shin2022) - Kim beta-range SSVEP: [Row in the open register →](https://bci.report/datasets/register/#r-site-kim2025-beta-range) - MIND: [Row in the open register →](https://bci.report/datasets/register/#r-site-mind2026) - eldBETA: [Row in the open register →](https://bci.report/datasets/register/#r-site-eldbeta) - Damm finger motor imagery: [Row in the open register →](https://bci.report/datasets/register/#r-site-damm-finger-imagery) - NeBULA: [Row in the open register →](https://bci.report/datasets/register/#r-site-nebula) - ZuCo 2.0: [Row in the open register →](https://bci.report/datasets/register/#r-site-zuco2) Methods & limits ## Methods and limits ### Shin2022: how the check was run Leave one participant out: 10 folds, each fitted on 9 participants. The window is the first second of feedback, read from its recorded start, never extended or padded: 250 samples at 250 Hz from a g.tec g.Nautilus RESEARCH (16 channel) headset with dry electrodes (EEG1–EEG16 in stored order; the headers carry no channel names). The label is the instructed target as each run stores it (TargetCode); the trial’s outcome is never read. ### Kim beta-range: how the baseline was run - **Algorithm**: Standard CCA, largest canonical correlation over the stored target frequencies; nothing is trained or tuned, and no trial is dropped. - **Channels**: 11 (OZ, O1, O2, POZ, PO3, PO4, PO7, PO8, PZ, P3, P4); the stored mean of M1 and M2 subtracted; no unit conversion. - **Window**: A nominal 2-second window from the stored onset, with no visual-delay shift: 2,048 samples at 1,024 Hz. - **Filter**: 13–44 Hz, second-order Butterworth band-pass, forward and backward, inside the window only. - **References**: Sine and cosine at each stored target frequency and its second harmonic. - **Metric**: Mean participant accuracy; with six trials per target for every participant it equals balanced accuracy in this cohort. ### MIND: how the pipeline was run - **Channels**: 64 EEG channels at 1,000 Hz; the auxiliary channels are left out. - **Filter**: FIR band-pass over each whole run, zero phase, 8–30 Hz; resampled to 200 Hz. - **Window**: After each stored class marker, 1,600 samples. - **Classifier**: CSP with 4 components and OAS regularisation, then shrinkage LDA on the four CSP log-power features; fitted on the training runs only, with no search. - **Not applied**: ICA, interpolation, rereferencing, amplitude clipping, channel search and trial exclusion. - **Ordinal lookup (no EEG)**: For each of the 20 positions of a pair within a run, the code the training runs hold most often at that position, ties to the lower code. - **Majority (no EEG)**: The code the training runs hold most often, ties to the lower code. ### eldBETA: how the two decoders were run - **Decoders**: Standard CCA with 3 harmonics, unfiltered, columns standardised; filter-bank CCA with 7 sub-bands whose lower edges run from 6 to 54 Hz up to 90 Hz and 5 harmonics, weights k^−1.25 + 0.25, each filter Chebyshev type I and zero-phase inside its window only; nothing is trained or tuned, and no trial is dropped. - **Channels**: 9 (PZ, PO5, PO3, POZ, PO4, PO6, O1, OZ, O2). - **Window**: 3 seconds, starting 140 samples after each event marker, at 1,000 Hz. - **Metric**: Mean participant accuracy; with seven trials per target for every participant it equals balanced accuracy in this cohort. ### The intervals Shin2022: each interval resamples the 10 participants 10,000 times, with one resampling index for all three arms, and reads its percentiles by linear interpolation. Kim beta-range: its interval resamples the 40 participants 20,000 times. eldBETA: its interval resamples the 100 participants 10,000 times, paired across the two decoders. ### What is not published **Shin2022, eldBETA and ZuCo 2.0:** - Participant rows of any kind: identifiers, run and file names, fold assignments, features, fitted models, decision values and predictions; raw signals from any of the three datasets. - Per-participant figures from the Shin2022 evidence: window counts per participant and their ranges, and which participants had flat windows. - The Shin2022 feedback-duration distribution (quantiles and mean) and the per-channel counts of the signal census. - eldBETA’s one-participant accuracy and its fixed first-frequency comparator. - ZuCo 2.0’s converted-run code, row-level timing values and marker counts; no ZuCo 2.0 data are redistributed. - MOABB’s catalog trial and run counts for Shin2022: register content, in the open register file. - Compute times, storage sizes, hosts and storage locations. - The research exports, audits, reviews, protocols and handoffs themselves: each is pinned by SHA-256 in the release manifest and checked by the export, never published, because several carry private storage paths. **eldBETA, all participants:** - Per-participant accuracies and correct counts, and any ordering of participants. - Participant identifiers of scored or excluded people. - The earlier one-participant score, which stays unrestated. - Per-trial predictions, correlations, sub-band values, per-block or per-trial file-check values, windows and signals. - Age, gender, BCIQ and SNR, which were never read. - Per-target accuracies, confusions and ceiling counts beyond the equal group. - Compute times, storage sizes, hosts and locations. - The research records, pinned by SHA-256. - No eldBETA data are redistributed. **Kim beta-range, MIND, Damm and NeBULA:** - Values per person, per record ID, per fold, per run or per trial: identifiers, file and run names, fold assignments, features, fitted models, decision values and predictions; raw signals from any of the four datasets. - Every Damm and NeBULA score: accuracies, balanced accuracies, control values, differences, paired correctness counts and diagnostic fractions; and any comparison of NeBULA’s order diagnostic with its EEG or EMG arms. - MIND’s pooled metrics, confusion matrices and macro-F1, the secondary panel’s schedule census and differences, the overfull runs’ marker counts and the model-fit counts. - Kim’s per-participant accuracies, the fixture identities of the independent audit and the stored-value ranges of the schema record. - Compute times, CPU and memory accounting, storage sizes, hosts and storage locations. - The research aggregates, audits, reviews, protocols and handoffs themselves: each is pinned by SHA-256 in the release manifest and checked by the export, never published, because several carry private storage paths. ### ZuCo 2.0: source and stated terms NEMAR accession nm000279, version 1.0.0, doi:10.82901/nemar.nm000279.v1.0.0. Licence as stated: CC-BY-4.0, in the dataset_description.json of that version. Not yet verified: the original terms on OSF; terms for the text excerpts the stimuli quote. [Dataset record ↗](https://data.nemar.org/nm000279/v1.0.0/) Hollenstein, N., Troendle, M., Zhang, C., & Langer, N. (2020). ZuCo 2.0: A Dataset of Physiological Recordings During Natural Reading and Annotation. LREC 2020. Data source: [data-quality-update.json](https://bci.report/data/data-quality-update.json) · schema bci-report-data-quality-update-v1 · [dataset-checks-update.json](https://bci.report/data/dataset-checks-update.json) · schema bci-report-dataset-checks-update-v1 · [eldbeta-cohort-update.json](https://bci.report/data/eldbeta-cohort-update.json) · schema bci-report-eldbeta-cohort-update-v1. ## Related - [How each source’s licence, consent and ethics statement was read →](https://bci.report/data-use/#sources-data-quality) - [How the licences and statements of Kim beta-range, MIND, Damm and NeBULA were read →](https://bci.report/data-use/#sources-dataset-checks) - Small pilot studies, each with its own limits: [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/) - [The open register: every EEG dataset we know of, and how far we checked it →](https://bci.report/datasets/register/) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *What do the files show once a public EEG dataset is checked end to end?* https://bci.report/topics/data-quality/ Figures from releases [`eldbeta-cohort-update-20261106`](https://bci.report/releases/#eldbeta-cohort-update-20261106) (2026-11-06), [`dataset-checks-update-20261028`](https://bci.report/releases/#dataset-checks-update-20261028) (2026-10-28), [`data-quality-update-20261020`](https://bci.report/releases/#data-quality-update-20261020) (2026-10-20). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/data-quality/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Do simple baselines work on signals other than scalp EEG? Other signals · human fNIRS and macaque intracortical spiking Three analyses: executed finger and foot tapping, classified from human fNIRS in people left out of training; finger velocity, decoded from one macaque’s motor-cortex spiking within one session; and a numerical check of every optical value in a second human fNIRS dataset. Two more sources are prepared, not yet scored. - [Reviewed aggregate JSON ↓](https://bci.report/data/beyond-scalp-update.json) - Measured on: [Bak et al. fNIRS finger- and foot-tapping dataset](https://bci.report/datasets/bak-fnirs/); [MC_RTT (Neural Latents Benchmark)](https://bci.report/datasets/nlb-mc-rtt/); [OpenNeuro ds007420 · multi-distance fNIRS](https://bci.report/datasets/ds007420/) ## Short answer Partly. On human fNIRS, a fixed shrinkage LDA told right-hand, left-hand and foot tapping apart in people it never saw: with each of **30** people left out in turn, its participant-macro balanced accuracy was **66.8%** (95% interval **60.6%** to **72.9%**), well above the training prior’s **33.3%**. These were executed movements, not motor imagery, and systemic or motion effects are not excluded. On one macaque’s motor-cortex spiking, a ridge decoder recovered part of the finger velocity on held-out trials of the same session: R² **0.240** against **−0.00175** for the training mean, and a lower root-mean-square error, but a higher mean absolute error (**0.038** against **0.033** m/s). On a second fNIRS dataset, a numerical coverage check finds all **48,502,600** optical values finite and positive. Human · fNIRS · executed movement · Bak et al. ## fNIRS: which movement did a person the classifier never saw make? Locally executed experiment · offline · leave one whole person out 30 people each tapped the fingers of the right hand, the fingers of the left hand or a foot when cued: executed movement, not motor imagery. That is 2,250 trials, 75 per person and 25 of each class. The signals are the supplied oxy- and deoxyhemoglobin series, 40 channels in mM·cm as delivered, not absolute concentrations. Each person is held out once: a fixed classifier is fitted on the other 29 people’s 2,175 trials and predicts all 75 of the held-out person’s, from the first 15 s after each task onset. The comparator is the training prior, which predicts the most frequent training class every time. **Participant-macro balanced accuracy for a person left out of training.** Dot: the mean over people, each weighted equally; line: the classifier’s descriptive 95% interval; dashed line: chance, one in three. - Shrinkage LDA — 95% interval 60.6%–72.9%: 66.8% - Training prior (comparator) — no interval: every person gets the same score: 33.3% Participant-macro balanced accuracy: each held-out person’s mean of the three class recalls, averaged over the 30 people with equal weight; chance is 33.3%. Class recall is computed over all held-out trials of every person together, the same number in each class: a secondary point estimate, without an interval. | Decoder | Balanced accuracy | Recall · right hand | Recall · left hand | Recall · foot | | --- | --- | --- | --- | --- | | Shrinkage LDA — standardization and shrinkage LDA, fitted on the training people only | 66.8% (95% interval 60.6%–72.9%) | 69.2% | 63.6% | 67.7% | | Training prior (comparator) — predicts right-hand tapping every time: the classes are balanced, and ties go to the fixed class order | 33.3% (no interval: every person gets the same score) | 100.0% | 0.0% | 0.0% | **66.8%** Shrinkage LDA, participant-macro balanced accuracy, 30 people each held out once; the training prior scores 33.3% ### Well above the training prior, in people the classifier never saw Its descriptive interval, 60.6% to 72.9%, lies wholly above chance. The interval resamples the fixed participant scores, 2,000 draws: it describes variation between people given the fitted folds, and because the folds share training people it is not a refit bootstrap and not an independent clinical confidence interval. The supplied series’ upstream preprocessing may use future samples, so the result does not apply to online control or a device’s response time. It does not establish a brain-only signal or cortical localization: task-correlated systemic or motion effects are not excluded, and the supplied files carry no probe coordinates. The protocol is this site’s own and does not reproduce the authors’ within-person scores. Data: Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong, `doi:10.6084/m9.figshare.9783755.v2`, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). [Dataset page →](https://bci.report/datasets/bak-fnirs/) · [Credit, licence and consent ↓](https://bci.report/topics/beyond-scalp-eeg/#credits) Independently audited: every input file hash, the marker-to-label order, whole-person folds, training-only scaling, the comparator’s tie rule and prediction alignment checked, and the class recalls, the mean and the interval recomputed without the frozen scoring code. The features were not regenerated from the raw channels, and the classifier was not refit. Macaque · intracortical spiking · MC_RTT (Neural Latents Benchmark) ## Spikes: how much finger velocity does a linear decoder recover within one session? Locally executed experiment · offline · one macaque, one session · not an official Neural Latents Benchmark score One rhesus macaque made self-paced reaches to targets on a grid while an electrode array in primary motor cortex recorded sorted units, in one session. A fixed ridge decoder reads the spike counts of the 98 held-in neurons (the benchmark’s 32 held-out neurons are not inputs) in 20 ms bins, with 160 ms of trailing history, and predicts the x/y finger velocity 140 ms later. It is fitted on the session’s stored 810 training trials and scored on its 270 validation trials, interleaved with them in time: 12,960 and 4,320 rows, none excluded. The comparator is the training mean: the training rows’ mean velocity, predicted for every row. **R² on the validation trials, for x, y and their mean: the ridge against the training mean.** Dot: the value; dashed line: an R² of zero. Higher is better. - Ridge on spike-count history · mean of x and y: 0.240 - Ridge on spike-count history · x: 0.202 - Ridge on spike-count history · y: 0.278 - Training mean (comparator) · mean of x and y: −0.00175 - Training mean (comparator) · x: −0.000104 - Training mean (comparator) · y: −0.00340 **RMSE on the validation trials, in m/s: lower is better.** Dot: the value; dashed line: the training mean’s RMSE. - Ridge on spike-count history: 0.056 - Training mean (comparator): 0.064 **MAE on the validation trials, in m/s: lower is better.** Dot: the value; dashed line: the training mean’s MAE. - Ridge on spike-count history: 0.038 - Training mean (comparator): 0.033 Every published metric on the 4,320 validation rows, with the arm that is better on it. R² is unitless; RMSE and MAE are in m/s. | Metric | Ridge on spike-count history | Training mean (comparator) | Better on this metric | | --- | --- | --- | --- | | R², mean of x and y — unitless; higher is better | 0.240 | −0.00175 | The ridge | | R², x — unitless; higher is better | 0.202 | −0.000104 | The ridge | | R², y — unitless; higher is better | 0.278 | −0.00340 | The ridge | | RMSE (m/s) — root-mean-square error; lower is better | 0.056 | 0.064 | The ridge | | MAE (m/s) — mean absolute error; lower is better | 0.038 | 0.033 | The training mean: the ridge’s error is larger | **0.240** R², mean of x and y, ridge on spike-count history; the training mean scores −0.00175 ### A lower squared error, but a higher absolute error The ridge lowers the root-mean-square error from 0.064 to 0.056 m/s, but its mean absolute error is higher, 0.038 against 0.033 m/s: it is not better on every metric. No interval is reported: the validation rows are temporally correlated samples from one animal in one session, not independent subjects. One macaque and one session, with training and validation trials interleaved: no cross-day, cross-animal or human evidence, and with one animal every figure is that animal’s. A supervised diagnostic on the stored trials, not co-smoothing and not a latent-model comparison. The velocity labels were derived offline from finger position, and averaging within each bin differs from the benchmark’s own resampling. Offline: not a real-time or closed-loop controller. Data: Joseph E. O’Doherty, Mariana M. B. Cardoso, Joseph G. Makin and Philip N. Sabes (Sabes Lab, UCSF), prepared for the Neural Latents Benchmark, `DANDI 000129 · 0.241017.1444`, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Animal data: collected under procedures approved by institutional animal care and use committees; informed consent does not apply. [Dataset page →](https://bci.report/datasets/nlb-mc-rtt/) · [Credit, licence and consent ↓](https://bci.report/topics/beyond-scalp-eeg/#credits) Independently audited: every feature and target row reconstructed independently of the code that produced it, the ridge coefficients checked against the normal equations, and the training-only normalization, prediction alignment and every metric recomputed. The stored velocity itself, which the dataset derives offline from finger position, was not checked. Human · fNIRS · ds007420 ## Optical recordings: are the values numerically usable? Locally executed check · label-independent · counts only 12 people, 60 recordings of resting state, left- and right-hand ball squeezing and instructed head and face movements, each with 200 optical channels: raw intensity at 760 and 850 nm for each source–detector pair. Every optical value of every recording was checked without labels: is it finite, is it positive, and is any channel exactly constant? Optical values checked 48,502,600 Every channel at each of the 242,513 time points of the 60 recordings: 12,000 recording-channels, no recording excluded. Finite and positive 48,502,600 Finite: 48,502,600, 100.0%. Positive: 100.0%. The denominator is every optical value. Zero · negative · non-finite 0 · 0 · 0 And 0 exactly constant channels: none whose samples are all finite and exactly equal, with no tolerance. Positive, non-constant values alone cannot establish signal quality: signal-to-noise, scalp coupling, saturation, motion, wavelength-pair integrity, physiological plausibility, neural origin, task validity and decoding are not assessed. The classifier is held, so no score exists. The numeric task codes have no verified lookup to hands or movements, and the event durations disagree: 1 s in the events table, 10 s in the optical files and 5 s in the cited paper. The 17 recordings held for their unverified auxiliary motion-sensor clocks stay held: optical coverage does not repair them. Data: Shibo Zhou, De’Ja Rogers, David A. Boas, Meryem A. Yücel and Yuanyuan Gao, `OpenNeuro ds007420 · 1.0.5`, [CC0](https://creativecommons.org/publicdomain/zero/1.0/). [Dataset page →](https://bci.report/datasets/ds007420/) · [Credit, licence and consent ↓](https://bci.report/topics/beyond-scalp-eeg/#credits) Independently audited: every channel count reconciled against the recording totals and the aggregate, and the code and its synthetic tests reviewed. The raw-signal pass was not repeated. Status only · no score ## Not yet evaluated Counts of what was prepared from the public files. ### FALCON H2 · human intracortical handwriting **Prepared, not scored.** Human, one participant · intracortical spiking · attempted handwriting of cued sentences, across recording days. `DANDI 000950 · 0.241029.1403`, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). No model has been fitted. The public release has 47 files over 26 recording dates, with 192 channels; its held-in calibration split has 671 trials. All 42 public minival trials were found inside those calibration trials, so a leakage-safe diagnostic must first remove them from training. That leaves 629 training trials, which carry 31 symbols (32 classes for a CTC decoder, with the blank). The 15 held-out calibration trials are inputs for few-shot recalibration, not independent evaluation data. The minival trials were inspected to find the overlap, so a later diagnostic on them would be exposed, not blinded, and would not be the official FALCON cross-day score. ### Rabut et al. · human functional ultrasound through a cranial window **Prepared, not analyzed.** Human, one participant · functional ultrasound through an acoustically transparent cranial window · drawing, guitar playing, and speech or reading, against rest. `doi:10.22002/f3y3k-em558`, [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Frame labels, timestamps and image axes are verified; no Doppler value has been read and no decoder has been fitted. Prepared: 9 runs of one person, not of nine participants — 7 drawing, 1 guitar and 1 mixed speech and reading — with 2,880 frames, 1,800 at rest and 1,080 during activity. Only rest against activity is verified; the speech and reading codes of the mixed run have no verified individual meaning. Measured timestamps are used, because the nominal block durations in the run labels disagree with the measured frame timing. Drawing, guitar and speech or reading are different paradigms and are not pooled into one task. The data are licensed for non-commercial use only. **ds007420 · fNIRS classifier · No score.** Its numeric task codes have no verified lookup to hands or movements, and its event durations disagree between the events table, the optical files and the cited paper, so no classifier was scored. Its optical coverage counts are published. [Its row in the holds register →](https://bci.report/releases/#hold-ds007420-classifier) Methods & limits ## Limits ### How each result was produced ### Human fNIRS, Bak et al. The supplied 40 hemodynamic series. For each trial, the channel mean in three consecutive 5 s windows from task onset to 15 s, each minus the mean of the second before onset: 120 features, with no filtering beyond the supplied series. Standardization and shrinkage LDA (lsqr solver, automatic shrinkage, empirical training priors) fitted on the training people only, with no channel, feature or hyperparameter search; 30 folds, each leaving one whole person out. ### Macaque spiking, MC_RTT Spike counts of the held-in neurons in 20 ms bins, smoothed over 3 trailing bins, with 6 history lags: 588 features, reset at every trial and observation gap. The target is the mean stored x/y finger velocity in a bin 140 ms ahead, 16 bins kept from each 600 ms trial. Standardization fitted on the training rows, then ridge regression (alpha 10, Cholesky solver, with intercept), training-constant features dropped, no hyperparameter search. Only the training file is used, with its stored training and validation trials; the test file was not opened. ### Human fNIRS coverage, ds007420 Every optical value of the 60 recordings read once, without labels or events. Shares use every optical value as the denominator; a constant channel is one whose samples are all finite and exactly equal, with no tolerance. ### What the interval covers Only the fNIRS result has an interval: a descriptive 95% percentile interval from 2,000 resamples, with replacement, of the 30 fixed participant scores. The MC_RTT figures have none, and the ds007420 figures are counts. ### Limits that hold for all three - Locally executed experiments, not scores reported in a paper or on a benchmark leaderboard. - The species, signals, tasks and metrics differ, so the three results share no scale and each is reported on its own; none is compared with a scalp-EEG result, and none joins the eight-protocol core matrix. - Offline only: no online or closed-loop control, device, clinical or deployment claim. - Fixed before scoring: no channel, feature or hyperparameter search, and no score here was used to choose a setting. ### Not published here Per-person, per-recording, per-channel, per-trial, per-fold and per-row values of any kind; the confusion matrices; participant, recording, file and session identifiers and recording dates; signals, features, predictions, model coefficients, normalization statistics, cue texts, labels, timestamps and images; measured compute; and demographics. Why so little is published for one animal and for small cohorts: [small cohorts, on the data-use page →](https://bci.report/data-use/#small-cohorts-pilots) ### Credits, licences, and consent or animal care ### Bak 2019 fNIRS dataset · unilateral finger and foot tapping Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong · Open Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, figshare, doi:10.6084/m9.figshare.9783755.v2, CC BY 4.0. Paper: Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong, Open-Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, Electronics 8(12), 1486 (2019), doi:10.3390/electronics8121486. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Approved by the Korea University Institutional Review Board (KUIRB-2019-0254-01); the experiments followed the criteria of the Declaration of Helsinki. The participants' written consent, approved by that board, was submitted before data recording. Read from an Internet Archive capture of the publisher’s article page, because the publisher refused automated requests. [Dataset record ↗](https://doi.org/10.6084/m9.figshare.9783755.v2) · [Paper ↗](https://doi.org/10.3390/electronics8121486) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Statement read from ↗](https://web.archive.org/web/20230731121639/https://www.mdpi.com/2079-9292/8/12/1486/htm) ### MC_RTT · macaque motor cortex during self-paced reaching (Neural Latents Benchmark) O'Doherty, Joseph (2024) MC_RTT: macaque motor cortex spiking activity during self-paced reaching (Version 0.241017.1444) [Data set]. DANDI archive. https://doi.org/10.48324/dandi.000129/0.241017.1444, CC BY 4.0. Recorded by Joseph E. O'Doherty, Mariana M. B. Cardoso, Joseph G. Makin and Philip N. Sabes in the Sabes Lab at UCSF (Nonhuman Primate Reaching with Multichannel Sensorimotor Cortex Electrophysiology, Zenodo, doi:10.5281/zenodo.3854034; procedure: Makin et al., Journal of Neural Engineering 15(2), 026010 (2018), doi:10.1088/1741-2552/aa9e95). Prepared for the Neural Latents Benchmark: Felix Pei, Joel Ye et al., Neural Latents Benchmark '21: Evaluating latent variable models of neural population activity, NeurIPS 2021 Track on Datasets and Benchmarks, arXiv:2109.04463. Derived analysis by BCI Report; not endorsed by the authors. **Animal care, as the Neural Latents Benchmark paper states it:** Collected under procedures and experiments approved by the Institutional Animal Care and Use Committees at the respective institutions; MC_RTT was collected in the Sabes Lab at UCSF. Informed consent: not applicable, animal data. [Dataset record ↗](https://doi.org/10.48324/dandi.000129/0.241017.1444) · [Paper ↗](https://arxiv.org/abs/2109.04463) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · statement read from the paper ### ds007420 · multi-distance fNIRS during ball squeezing and motion-artifact tasks Shibo Zhou, De’Ja Rogers, David A. Boas, Meryem A. Yücel and Yuanyuan Gao · Open-access Multi-Distance fNIRS Dataset Generated During Ball-Squeezing and Motion-Artifact Induction Tasks, OpenNeuro, doi:10.18112/openneuro.ds007420.v1.0.5, CC0. The dataset asks to acknowledge: Gao, Y., Rogers, D., von Lühmann, A., Ortega-Martinez, A., Boas, D.A., & Yücel, M.A. (2023). Short-separation regression incorporated diffuse optical tomography image reconstruction modeling for high-density functional near-infrared spectroscopy. Neurophotonics, 10(2), 025007. https://doi.org/10.1117/1.NPh.10.2.025007. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Approved by the Institutional Review Board of Boston University (Protocol #4502, as the dataset description records it). All participants provided written informed consent to take part in the study. [Dataset record ↗](https://doi.org/10.18112/openneuro.ds007420.v1.0.5) · [Paper ↗](https://doi.org/10.1117/1.NPh.10.2.025007) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) · [Statement read from ↗](https://europepmc.org/article/PMC/PMC10203730) ### FALCON H2 · human intracortical handwriting Fan, Chaofei; Hahn, Nick; Kamdar, Foram; Avansino, Donald; Wilson, Guy; Hochberg, Leigh; Shenoy, Krishna V; Henderson, Jaime; Willett, Frank (2024) FALCON Benchmark H2: Human Handwriting iBCI (Version 0.241029.1403) [Data set]. DANDI archive. https://doi.org/10.48324/dandi.000950/0.241029.1403, CC BY 4.0. Described in: Plug-and-Play Stability for Intracortical Brain-Computer Interfaces: A One-Year Demonstration of Seamless Brain-to-Text Communication, NeurIPS 2023. Part of the FALCON benchmark. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Permission from the Institutional Review Boards of Stanford University (protocol #20804); the BrainGate2 pilot trial ran under FDA Investigational Device Exemption #G090003 (ClinicalTrials.gov NCT00912041). The participant gave informed consent and was enrolled in the BrainGate2 trial. [Dataset record ↗](https://doi.org/10.48324/dandi.000950/0.241029.1403) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Statement read from ↗](https://api.dandiarchive.org/api/dandisets/000950/versions/0.241029.1403/) ### Rabut et al. human functional ultrasound through a cranial window Claire Rabut, Sumner Norman, Whitney Griggs, Kay Jann, Jonathan Russin, Vasileios Christopoulos, Charles Liu, Richard A. Andersen and Mikhail G. Shapiro · Dataset for: A window to the brain: ultrasound imaging of human neural activity through an acoustically transparent cranial prosthetic, CaltechData, doi:10.22002/f3y3k-em558, CC BY-NC 4.0. Paper: Claire Rabut, Sumner L. Norman, Whitney S. Griggs, Jonathan J. Russin, Kay Jann, Vasileios Christopoulos, Charles Liu, Richard A. Andersen and Mikhail G. Shapiro, Functional ultrasound imaging of human brain activity through an acoustically transparent cranial window, Science Translational Medicine 16, eadj3143 (2024), doi:10.1126/scitranslmed.adj3143. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Approved by the Institutional Review Boards of the University of Southern California, Caltech (IR19-0902) and Rancho Los Amigos National Rehabilitation Hospital. The participant's consent was obtained. [Dataset record ↗](https://doi.org/10.22002/f3y3k-em558) · [Paper ↗](https://doi.org/10.1126/scitranslmed.adj3143) · [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) · [Statement read from ↗](https://www.vis.caltech.edu/documents/28341/scitranslmed.adj3143.pdf) Derived analyses by BCI Report, aggregates only; not endorsed by the data’s authors, and no raw data are redistributed. Data source: [beyond-scalp-update.json](https://bci.report/data/beyond-scalp-update.json) · schema bci-report-beyond-scalp-update-v1. ## Related questions - Small EEG pilots, some on datasets that also record fNIRS or muscle activity: [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/) - Electrodes in the ear rather than on the scalp, against a full scalp montage: [Can fewer electrodes, or electrodes in the ear, match a full scalp montage?](https://bci.report/topics/fewer-electrodes/) - Continuous cursor tracking decoded from scalp EEG across sessions of the same person, a different signal, task and metric: [Continuous cursor tracking, a negative result](https://bci.report/topics/later-sessions/#forenzo). - EEG and EMG recorded together while one participant reaches with an upper-limb exoskeleton, checked from its files, with its scores held: [NeBULA, a status case on the data-quality page](https://bci.report/topics/data-quality/#nebula). Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *Do simple baselines work on signals other than scalp EEG?* https://bci.report/topics/beyond-scalp-eeg/ Figures from release [`beyond-scalp-update-20261012`](https://bci.report/releases/#beyond-scalp-update-20261012) (2026-10-12). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/beyond-scalp-eeg/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # What can a small pilot show, and what can it not? Pilot studies · small cohorts Four small pilots: injected orthosis errors in one cohort of eight people, including frozen EEGPT against matched classical features; inner speech in Greek and English, across people and languages; channels flat at a digital rail in an imagined-handwriting dataset, and what the EEG left over can show; and a stored-clip check on an emotion dataset. - [Reviewed aggregate JSON ↓](https://bci.report/data/pilot-studies-update.json) - Measured on: [Orthosis error-detection dataset (Kueper et al. 2024)](https://bci.report/datasets/orthosis-error-detection/); [BINGO cross-linguistic inner-speech dataset](https://bci.report/datasets/bingo-imagined-speech/); [EEG-fNIRS-based Handwriting Trajectory Dataset](https://bci.report/datasets/eeg-fnirs-handwriting/); [REFED emotion dataset (EEG-fNIRS)](https://bci.report/datasets/refed/) ## Short answer Five findings from four pilots. Orthosis device errors: fixed EEG features reach **73.0%** balanced accuracy in a person left out of training (95% interval **68.0%** to **77.7%**), against **51.4%** for a confound probe; the labels are recorded device events, so movement or device artifacts may contribute. Frozen EEGPT against matched classical features on the same eight people: **−0.6 pp** (interval **−4.0** to **+3.2 pp**): the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. Inner speech in Greek and English (BINGO): every direction stays near the **6.7%** prior, within and across languages. Handwriting, **5** people: with flat and rail-touching channels masked, relative EEG reaches **26.81%** balanced accuracy on four coded labels against **25.00%** for the majority control, above the missingness-mask controls but not separated from elapsed time and trial order: a data-quality case study, not handwriting recognition. REFED emotion clips: EEG band power has a larger error than every control that uses no EEG, and when videos are shared, looking up the same video’s rating has the smallest. Human · scalp EEG · recorded orthosis device events · Kueper et al. ## Orthosis errors: can EEG flag an injected device error in a person left out of training? Locally executed experiment · offline · one cohort in three configurations, counted once 8 people wore an active orthosis that moved the right arm in elbow flexion and extension, and errors were injected into some of its movements. The task tells the recorded error events from standard ones, in the 0–0.5 s after each event: 2,407 events in 83 runs, 474 errors and 1,933 standard. Each person is held out whole, and a fixed probe fitted on the others scores them; results are averaged over the 8 people with equal weight. ### The 64-channel classical baseline **Balanced accuracy for a person left out of training, 64-channel configuration.** Dot: the mean over the eight people, each weighted equally; line: the 95% person-bootstrap interval; dashed line: chance. The permutation control is drawn at the mean of its replicates; the range in its label is not an interval. - Fixed EEG features + logistic regression, 64 channels — 95% interval 68.0%–77.7%: 73.0% - Direction/time confound probe (no EEG) — 95% interval 50.4%–52.2%: 51.4% - Training class prior — a constant probability: chance for every person: 50.0% - Label-permutation control — mean of 20 replicates, which ranged from 48.0% to 52.6%: 50.1% **73.0%** Fixed EEG features + logistic regression, 64 channels: balanced accuracy for a person left out of training. The confound probe scores 51.4%, the class prior 50.0%. ### Above every control, in people the probe was not trained on Its 95% interval, 68.0% to 77.7%, lies above the confound probe’s interval, the class prior and every permutation replicate. Matched person by person, EEG minus the confound probe is +21.6 pp, with no interval computed; EEG was above the probe for 8 of 8 people. The labels are recorded device events, and the result does not show that a person recognised the error: movement, muscle and device-injection artifacts may contribute, and beating a probe of movement direction and timing does not isolate a neural origin. This reanalysis does not reproduce the authors’ competition score. ### Frozen EEGPT against the same baseline on 58 channels EEGPT supports 58 of the 64 channels; the other six (TP9, TP10, PO9, PO10, FT9, FT10) are left out, fixed before scoring. On those channels the same classical baseline scores 71.0% (95% interval 64.8% to 76.9%), also above its confound probe, the prior and every permutation replicate: a second configuration of the same people, not a replication, and no conclusion about how many channels are needed follows. Frozen EEGPT then replaces the classical features and feeds the same probe, with the same people, folds and permutations. **Balanced accuracy for a person left out of training, 58-channel configuration: frozen EEGPT against the matched classical features.** Dot: the mean over the eight people; line: the 95% person-bootstrap interval; dashed line: chance. - Frozen EEGPT + the same probe, 58 channels — 95% interval 66.2%–73.9%: 70.3% - Fixed EEG features + logistic regression, 58 channels — 95% interval 64.8%–76.9%: 71.0% - Direction/time confound probe (no EEG) — 95% interval 50.4%–52.3%: 51.4% - Training class prior — a constant probability: chance for every person: 50.0% **The paired difference, frozen EEGPT minus the matched classical features, in percentage points.** Dot: the mean of the eight people’s differences; line: its 95% person-bootstrap interval, which resamples each person’s difference; dashed line: no difference. - Frozen EEGPT minus matched classical features — 95% interval −4.0 to +3.2 pp: −0.6 pp **−0.6 pp** Frozen EEGPT minus the matched classical features, balanced accuracy, paired over the same 8 people; 95% interval −4.0 to +3.2 pp ### The paired interval spans zero Frozen EEGPT scores 70.3% (66.2% to 73.9%) and the matched classical features 71.0%: neither arm is shown to be better, and the two are not shown to be equivalent. One short-window setup on eight people is not a verdict on EEG foundation models in general. The same eight people and the same recorded device events as the 64-channel result: not additional coverage, and movement, muscle and device artifacts may contribute here too. The 64-channel scores were seen before this comparison was frozen, so it is a prospective diagnostic, not a blind confirmation. The probe is trained on the other seven people’s labels: supervised transfer, not zero-shot. Overlap between this dataset and EEGPT’s pretraining data has not been excluded, and the montage, filtering, reference and half-second window differ from its pretraining. Every published metric of the three configurations, each the mean over the 8 people with its 95% interval where it has one. Errors are 19.7% of the events, so average precision is read against a constant guess’s 0.197. AUROC and average precision are secondary metrics. | Arm | Balanced accuracy | AUROC | Average precision | Label-permutation control | | --- | --- | --- | --- | --- | | 64-channel configuration | | | | | | Fixed EEG features + logistic regression, 64 channels | 73.0% (95% interval 68.0%–77.7%) | 0.812 (95% interval 0.751–0.863) | 0.585 (95% interval 0.492–0.672) | 50.1% (range 48.0% to 52.6%) | | Direction/time confound probe (no EEG) | 51.4% (95% interval 50.4%–52.2%) | 0.524 (95% interval 0.507–0.541) | 0.228 (95% interval 0.208–0.249) | — | | Training majority | 50.0% (the same for every person) | 0.500 (the same for every person) | 0.197 (95% interval 0.194–0.200) | — | | Training class prior | 50.0% (the same for every person) | 0.500 (the same for every person) | 0.197 (95% interval 0.194–0.200) | — | | 58-channel configuration · the same people and folds | | | | | | Fixed EEG features + logistic regression, 58 channels | 71.0% (95% interval 64.8%–76.9%) | 0.793 (95% interval 0.719–0.854) | 0.544 (95% interval 0.438–0.654) | 50.1% (range 48.2% to 51.7%) | | Frozen EEGPT + the same probe, 58 channels | 70.3% (95% interval 66.2%–73.9%) | 0.780 (95% interval 0.721–0.827) | 0.564 (95% interval 0.489–0.630) | 49.8% (range 47.6% to 52.2%) | | Direction/time confound probe (no EEG) | 51.4% (95% interval 50.4%–52.3%) | 0.524 (95% interval 0.506–0.541) | 0.228 (95% interval 0.208–0.249) | — | | Training majority | 50.0% (the same for every person) | 0.500 (the same for every person) | 0.197 (95% interval 0.194–0.200) | — | | Training class prior | 50.0% (the same for every person) | 0.500 (the same for every person) | 0.197 (95% interval 0.194–0.200) | — | A response-clean subset is published as counts only: 451 events, 446 errors and 5 standard events, with 3 people having both classes. Its scores stay out of every comparison. Data: Niklas Kueper, Kartik Chari, Judith Bütefür, Julia Habenicht, Tobias Rossol, Su Kyoung Kim, Marc Tabie, Frank Kirchner and Elsa Andrea Kirchner, `doi:10.5281/zenodo.8345429`, [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). [Dataset page →](https://bci.report/datasets/orthosis-error-detection/) · [Credit, licence and consent ↓](https://bci.report/topics/pilot-studies/#credits) Independently audited: every score and interval of the three configurations recomputed from the sealed predictions, and for the comparison, refits of the main and confound probes reproduced the sealed probabilities exactly. The encoder features were not re-extracted, and not every permutation probe was refitted. Human · scalp EEG · inner speech in Greek and English · BINGO project ## Inner speech: can imagined words be decoded across people and across languages? Locally executed experiment · offline · a subject field left out at a time · a negative result 17 subject fields, as the source defines them (not proven to be distinct people), each said 15 matched nouns silently, in Greek and in English, 10 times per noun and language: 5,100 trials. A fixed spectral baseline — relative band power from 19 channels at 300 Hz, and a ridge classifier — is trained on one language of the other fields (2,400 trials) and tested on the held-out field in the same language and in the other, from the 1.5 s after the fixation cross. Both languages of the held-out field stay out of every fit. Chance is the prior of one in fifteen, 6.7%. **Balanced accuracy for a subject field left out of training, by training and test language.** Dot: the mean over the held-out fields, each weighted equally; line: its approximate descriptive 95% interval; dashed line: the prior, one in fifteen. Rust: within a language; green: across languages. - Greek → Greek — within a language · 95% interval 5.8%–6.9%: 6.4% - Greek → English — across languages · 95% interval 5.5%–7.6%: 6.6% - English → English — within a language · 95% interval 5.8%–7.6%: 6.7% - English → Greek — across languages · 95% interval 6.2%–7.9%: 7.1% Means over the 17 held-out fields, each weighted equally, with approximate descriptive 95% intervals. The prior is 6.7% for balanced accuracy and 20.0% for top-3 accuracy. The shuffle column gives the mean of 20 training-label shuffles and the lowest and highest replicate means. | Training → test language | Balanced accuracy | Top-3 accuracy | Order-only control | High-frequency ratio control | Training-label shuffles | | --- | --- | --- | --- | --- | --- | | Greek → Greek — within a language | 6.4% (95% interval 5.8%–6.9%) | 19.1% (95% interval 18.0%–20.0%) | 7.1% (95% interval 6.4%–7.9%) | 7.1% (95% interval 6.4%–7.8%) | 6.6% (range 5.5% to 7.5%) | | Greek → English — across languages | 6.6% (95% interval 5.5%–7.6%) | 21.0% (95% interval 19.6%–22.4%) | 6.2% (95% interval 5.4%–7.0%) | 6.4% (95% interval 5.7%–7.1%) | 6.5% (range 5.4% to 7.3%) | | English → English — within a language | 6.7% (95% interval 5.8%–7.6%) | 19.9% (95% interval 18.2%–21.6%) | 7.4% (95% interval 6.7%–8.1%) | 6.5% (95% interval 5.6%–7.5%) | 6.8% (range 6.1% to 7.8%) | | English → Greek — across languages | 7.1% (95% interval 6.2%–7.9%) | 19.6% (95% interval 18.3%–20.9%) | 6.5% (95% interval 5.8%–7.3%) | 7.2% (95% interval 6.5%–8.0%) | 6.8% (range 5.5% to 7.7%) | **Across languages minus within, balanced accuracy, in percentage points.** Dot: the mean over the held-out fields; line: its approximate descriptive 95% interval; dashed line: no difference. - Mean over both training languages — 95% interval −0.6 to +1.3 pp: +0.3 pp - Trained on Greek: English minus Greek — 95% interval −1.3 to +1.7 pp: +0.2 pp - Trained on English: Greek minus English — 95% interval −0.9 to +1.7 pp: +0.4 pp **+0.3 pp** Across languages minus within, balanced accuracy, the mean over both training languages; 95% interval −0.6 to +1.3 pp ### Near the prior in all four directions Balanced accuracy runs from 6.4% (Greek → Greek) to 7.1% (English → Greek), and every interval holds the prior of 6.7%; top-3 accuracy runs from 19.1% to 21.0%, against its prior of 20.0%. The contrast’s interval spans zero: no superiority or equivalence between languages, and no significance test. A negative result for one fixed baseline in this setting: it does not show that inner speech cannot be decoded by other methods. **Delivered signals, from the project’s deliverable [D3.1 – The BINGO Benchmarking Framework, version 2.0 ↗](https://bingo-project.gr/download/397):** The project documents the same preprocessing for both of its released datasets: band-pass filtering from 1 to 145 Hz, notch filtering at 50 Hz and its harmonics, Artifact Subspace Reconstruction (ASR), FORCe removal of ocular and muscle artifacts, and then segmentation into trials. The signals analysed here were therefore artifact-cleaned on the continuous recordings before any trial was cut. As the project states it; not re-verified against the delivered file. The subject fields are the source’s operational groups; whether they are distinct people, and how recordings group across days, was not independently certified. Physical amplitude units, the named montage and the display latency are not established, and the window, though verified against the stored time vectors, does not certify causal or online processing or a brain-only origin of the signal. No foundation model or deep model was evaluated. **Reported by the BINGO project, not run here.** On its other dataset, NATO-alphabet inner speech with 26 words, the project’s own leave-one-subject-out evaluation of EEGNet reports a mean accuracy of 3.85% against a chance level of 3.85%: at chance for people the model was not trained on. A different dataset, model and metric, so it is not comparable with the results above. [D3.1 – The BINGO Benchmarking Framework, version 3.0 ↗](https://bingo-project.gr/download/402) Data: the BINGO (Brain Imagined-Speech Communication) project, CERTH, [datasets page ↗](https://bingo-project.gr/datasets/) · documentation: [D3.1 – The BINGO Benchmarking Framework, version 2.0 ↗](https://bingo-project.gr/download/397). No licence is stated by the source. [Dataset page →](https://bci.report/datasets/bingo-imagined-speech/) · [Credit, licence and consent ↓](https://bci.report/topics/pilot-studies/#credits) Independently audited: every sealed prediction, aggregate, interval and contrast checked; the main and control models refitted with no difference, and the shuffle models reconstructed rather than refitted. The raw file was not re-extracted: the audit used cached features with verified lineage. Human · scalp EEG · imagined handwriting, four released coded labels · competition dataset ## Handwriting: how much of a recording with flat or rail-touching channels is usable? Locally executed experiment · offline · a person left out at a time · a data-quality and methods case study 5 people from the competition’s training split, with all 45 of their sessions: 1,620 events, of which 1,618 are scored; 2 events whose labels contradict each other stay quarantined. Each trial carries one of 4 released coded labels; which character or trajectory each label stands for is not established, so this is not handwriting recognition. Many of the recorded channels sit flat at the lower digital rail declared in the file header. The mask-aware analysis scores each trial only on the channels that pass a fixed availability rule, and sets that EEG against three controls that see no EEG: which channels are missing, the elapsed time and trial order, and both together. Each person is held out whole and scored by a model fitted on the other four; results are averaged over the 5 people with equal weight. **Balanced accuracy for a person left out of training, four coded labels.** Dot: the mean over the five people, each weighted equally; line: the descriptive 95% person-bootstrap interval; dashed line: the training-majority control, which scores chance for every person. The mask-only control has no interval: with five people, its bounds would show individual people’s scores. - Mask-aware relative EEG — 95% interval 26.30%–27.20%: 26.81% - Missingness mask only — no EEG · interval not published: 24.99% - Elapsed time / trial order — no EEG · 95% interval 24.03%–26.79%: 25.44% - Mask + time — no EEG · 95% interval 23.90%–26.12%: 25.01% - Training majority — chance for every person: 25.00% **The paired difference, mask-aware relative EEG minus each control, in percentage points.** Dot: the mean of the five people’s differences; line: its descriptive 95% person-bootstrap interval, which resamples each person’s difference; dashed line: no difference. - Minus the missingness mask — 95% interval +1.30 to +2.21 pp: +1.82 pp - Minus elapsed time / trial order — 95% interval −0.12 to +3.07 pp · crosses zero: +1.37 pp - Minus mask + time — 95% interval +0.43 to +3.05 pp: +1.80 pp Means over the 5 held-out people, each weighted equally, with descriptive 95% intervals where they are published. The shuffle column gives the mean of 5 training-label shuffles, which keep each person’s and session’s labels together, and the lowest and highest replicate means: a sensitivity check, not a significance test. | Arm | Balanced accuracy | Training-label shuffles | | --- | --- | --- | | Mask-aware relative EEG | 26.81% (95% interval 26.30%–27.20%) | 25.20% (range 23.85% to 26.69%) | | Missingness mask only — no EEG | 24.99% (interval not published) | 24.95% (range 24.88% to 25.00%) | | Elapsed time / trial order — no EEG | 25.44% (95% interval 24.03%–26.79%) | 24.70% (range 24.30% to 25.16%) | | Mask + time — no EEG | 25.01% (95% interval 23.90%–26.12%) | 24.71% (range 24.11% to 25.78%) | | Training majority — no EEG | 25.00% (chance for every person) | — | **26.81%** Mask-aware relative EEG: balanced accuracy for a person left out of training. The training-majority control scores 25.00%. ### Just above chance, and not separated from elapsed time and trial order Its interval, 26.30% to 27.20%, lies above the majority control. Paired person by person, relative EEG minus the missingness mask is +1.82 pp (+1.30 to +2.21 pp) and minus mask + time +1.80 pp (+0.43 to +3.05 pp); minus elapsed time and trial order it is +1.37 pp, and that interval, −0.12 to +3.07 pp, crosses zero. ### What the masks leave Every scored trial keeps between 27 and 38 of the 59 recorded EEG channels under the fixed availability rule — not constant, not touching either digital rail, numerically valid — so the recordings remain partly usable. “Available” is that numerical rule, not certified clean EEG; why the other channels sit at the rail is not known. Across 95,462 channel-trials, 50,811 carry the constant flag and 50,815 the lower-rail flag; the two flags overlap, so they are not summed. No upper-rail, non-finite or invalid-denominator flag occurred. These counts describe the five people analysed, not the whole dataset. Scored trials by how many of the 59 recorded EEG channels are available. | Available channels | Trials | | --- | --- | | 27 | 1,429 | | 28 | 65 | | 29 | 16 | | 30 | 36 | | 37 | 36 | | 38 | 36 | An earlier analysis of the same 5 people and 1,618 trials, with every channel on a common-average reference and log band power, scored 27.05% with logistic regression and 27.42% with shrinkage LDA. Its checks found a sample at a digital rail in 1,618 of 1,618 windows, and in 5 windows inspected afterwards, 148 of 295 channel-windows flat at the lower rail. It is the context for the mask-aware analysis, not a second cohort. Masking, referencing and the spectral representation changed together, so the difference between the two analyses cannot be attributed to masking, and no best model is chosen from them. These five people were the development cohort of the earlier analysis, whose scores were seen before this design: a quality-informed diagnostic, not independent confirmation, and no population inference. The official ten-person test data were not used. The optical species, units and the clock between the two modalities are unresolved, so no fNIRS or fusion conclusion follows, and there is no online, physiological-origin, competition-standing or medical claim. Data: [lasfk on Hugging Face · EEG-fNIRS-based Handwriting Trajectory Dataset ↗](https://huggingface.co/datasets/lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset), `revision 4036680b5fb0015472d55589da2f34c5a87a3c10`, [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) as declared on the dataset card, gated. No ethics-approval or informed-consent statement was found; the aggregates are published with that gap stated, by the owner’s decision. [Dataset page →](https://bci.report/datasets/eeg-fnirs-handwriting/) · [Credit, licence and consent ↓](https://bci.report/topics/pilot-studies/#credits) Independently audited: all 120 stored models reconstructed and the 20 true-label models refitted, with no discrepancy. The shuffle models were not refitted, the raw samples were not reread and the features were not re-extracted. Human · scalp EEG, stored clips · self-reported emotion · REFED ## Emotion ratings: what does a stimulus-identity control reveal? Locally executed computational diagnostic · offline · two nominal folders 2 nominal participant folders watched the same 15 videos: 30 stored EEG clips, not as many independent people. For each clip the target is the mean of its stored valence and arousal ratings on the published raw scale from 1 to 255: descriptive self-report means, not SAM scores. Four fixed arms are scored by equal-axis mean absolute error, lower being better: EEG Ridge on relative band power at every channel position, and three controls that use no EEG — the training clips’ mean, the training target of the same video, and a ridge on stored duration and the numeric video ID. Two conditions: a folder held out with the videos shared (2 folds), and a folder held out together with 3 of the videos it is tested on (10 folds). **Equal-axis mean absolute error with the videos shared and a folder held out, raw label units: lower is better.** Dot: the error over all clips of the condition; dashed line: the training-set mean’s error. - EEG Ridge (relative band power): 107.182 - Training-set mean — uses no EEG: 40.287 - Same-video training lookup — uses no EEG: 27.325 - Duration / numeric-video-ID Ridge — uses no EEG: 37.910 **Equal-axis mean absolute error with a folder and its test videos held out together, raw label units: lower is better.** Dot: the error over all clips of the condition; dashed line: the training-set mean’s error. - EEG Ridge (relative band power): 91.984 - Training-set mean — uses no EEG: 41.032 - Same-video training lookup — uses no EEG: 41.032 - Duration / numeric-video-ID Ridge — uses no EEG: 47.098 Equal-axis mean absolute error over the 30 clips of each condition, in raw label units; lower is better. In the second condition the lookup finds a matching video for 0 of 30 clips, so it falls back to the training-set mean. | Arm | Shared videos, a folder held out | Folder and videos held out together | | --- | --- | --- | | EEG Ridge (relative band power) | 107.182 | 91.984 | | Training-set mean — uses no EEG | 40.287 | 41.032 | | Same-video training lookup — uses no EEG | 27.325 | 41.032 | | Duration / numeric-video-ID Ridge — uses no EEG | 37.910 | 47.098 | **107.182** EEG Ridge, equal-axis mean absolute error with the videos shared; the same-video lookup scores 27.325, the training-set mean 40.287 ### EEG band power has the largest error in both conditions Held out together with the videos, EEG Ridge’s error is 91.984 against 41.032 for the training-set mean. With shared videos, looking up the same video’s training target has the smallest error: a familiar stimulus carries the prediction through its identity. This illustrates the confound; it does not establish a mechanism. The clips come from two nominal folders sharing their video keys, so no interval or significance test is reported, and no population, cross-day or physiological claim follows. The rating arrays had been inspected and their means computed before this design, so these are not untouched targets. Stored EEG units, the EEG-to-rating clock and the crop end points are unresolved: published video durations exceed the stored arrays by two seconds, and pairing whole arrays establishes no precise alignment. The numeric video ID is a key index, not an established presentation order. Data: Xiaojun Ning, Jing Wang, Zhiyang Feng, Tianzuo Xin, Shuo Zhang, Shaoqi Zhang, Zheng Lian, Yi Ding, Youfang Lin and Ziyu Jia, `REFED2025/REFED-dataset · revision 0298355c2823946ac5800b2ec12a082bf2dcd01b`, [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/), gated. [Dataset page →](https://bci.report/datasets/refed/) · [Credit, licence and consent ↓](https://bci.report/topics/pilot-studies/#credits) Independently audited: the sealed predictions reconstructed exactly, every metric recomputed independently and every preparation record verified. No raw signal or original annotation was reread, and no model was refitted. Status only · no figure ## Not yet evaluated These analyses lack the evidence a valid evaluation needs. ### Orthosis error-detection dataset: EMG and paired EEG–EMG analyses **No EMG, TinyMyo or EEG–EMG fusion result exists for the Orthosis dataset.** The exact clock anchors between the EMG sample rows and the EEG events, the latency and end points, and the meaning and units of the exported columns are not established. A nominal sampling rate alone does not establish the labels. ### EEG-fNIRS handwriting dataset: optical and fused analyses **No fNIRS or EEG–fNIRS fusion result exists for the handwriting dataset.** The optical species and column meanings, their units and a verified clock between the two modalities are not established. ### REFED: physiological, pointwise and fused analyses **No physiological or precisely aligned fusion evaluation exists for REFED beyond the stored-clip computational diagnostic.** Physical calibration, the crop boundaries, the mapping of the six optical components and paired timing are not established. ### Synchronized multimodal meditation and rest dataset (Zenodo): EEG phase labels and paired analyses **No meditation EEG phase prediction or paired analysis exists.** An explicit association between events and records, and clock anchors, are missing. One inspected record has a constant marker column, and that finding is not generalised to the other records. In this dataset RR means heartbeat RR intervals, not respiratory rate. Methods & limits ## Limits ### How each result was produced ### Orthosis, classical features The 64 EEG channels at 500 Hz with the declared header scaling, the recorded reference kept and the acceleration channels left out; continuous forward-only 1–30 Hz filtering that uses no future samples; the complete 0–0.5 s window (250 samples) after each event. 576 fixed amplitude and log band-power values per event (522 on the 58 channels), a training-only standardization and a balanced L2 logistic regression with C = 1, no tuning. Main-task runs only: 8 baseline runs, 82 targets without a following movement and 1 window crossing a boundary are left out, none depending on whether the person responded correctly. Every prediction was sealed before a separate scorer read the held-out labels. ### Orthosis, frozen EEGPT The frozen 25,287,168-parameter EEGPT target encoder at its native input of 58 channels by 128 samples: the same filtered window, resampled from 500 to 256 Hz, microvolts divided by 1,000; its summary tokens, averaged over two temporal patches, feed the same probe. The encoder’s weights are unchanged: no fine-tuning and no LoRA. Across the 8 folds, 344 probes were fitted. ### BINGO, inner speech Per trial and channel, max-absolute scaling and demeaning, then Hann periodograms; four relative bands partition 1–30 Hz across the 19 channels in source order, 76 values per trial, with no voltage calibration, montage interpolation or re-referencing. A source-only standardization and a ridge classifier with fixed alpha 1. Controls: an order-only control, a high-frequency ratio control, the class prior and 20 training-label shuffles per direction; all 816 prediction checkpoints were sealed before scoring. One incomplete field is left out by a complete-record rule set before scoring. The protocol, from the project’s D3.1, version 2.0: the word is shown at −2.5 s, removed for a countdown at −1.0 s, and the fixation cross at 0.0 s starts inner speech until 1.5 s; 15 nouns, 10 repetitions each, in fully randomized order. ### Handwriting, mask-aware EEG A fixed window 0.5–4.5 seconds after the released EEG-relative onset (4,000 samples at 1,000 Hz), read from the 59 header-identified EEG channels only; the EOG, HECG and annotation channels and the optical samples are left out. Labels are the released event codes 200 to 203, mapped to 0 to 3. Before any transform, a channel-window is masked if it is constant, touches either header-declared digital rail or is not finite; the available channels are demeaned one by one, with no common-average reference and no interpolation. Welch spectra (1,000-sample Hann window, 500 overlap) give five half-open bands from 1 to 45 Hz, each divided by the sum of the five: 295 channel-major values per trial, zero where a channel is unavailable. Source-only standardization and a balanced multinomial logistic regression with C = 1, no tuning. The controls fit the same model on the availability bits and the available-channel fraction, on four elapsed-time and trial-order features, and on both. All 125 checkpoints, for 120 model fits and five majority controls, were sealed before a separate scorer read the held-out labels. ### REFED, stored clips Five relative Welch band powers between 1 and 45 Hz, each divided by the channel’s total, at every channel position, with no channel selection or re-referencing; source-only standardization and Ridge with fixed alpha 1, no tuning, and predictions left unclipped. Every clip passed the numerical checks (1,920 channel–clip rows, 0 removed): finite, non-constant values and valid spectral denominators, which is numerical eligibility, not signal quality. ### What the intervals cover Orthosis: pointwise 95% percentile intervals from 10,000 person-level bootstrap draws, conditional on the sealed predictions; with eight people they are wide, and the paired interval resamples each person’s difference rather than comparing two separate intervals. BINGO: approximate descriptive intervals from 10,000 field bootstrap draws over overlapping leave-one-field-out fits; they omit refit variability and are not calibrated population confidence intervals. Handwriting: descriptive 2.5–97.5% intervals from 10,000 fixed-seed person-bootstrap draws over the five people, conditional on the already fitted overlapping folds; they are not calibrated population confidence intervals and not evidence of superiority. Each bound is a percentile of resampled five-person means, not one person’s score. REFED: none, because two nominal folders support no population inference. The permutation and shuffle controls are diagnostics, not significance tests. ### Limits that hold for all four - Each pilot answers its own question on its own dataset, task and metric; no result is set against another or against the core matrix. - The three Orthosis results are one cohort, counted once; no figure combines datasets, configurations, languages or conditions. - Reanalyses executed locally by this site, not results reported by the datasets’ authors. - Offline only: no online, real-time, clinical or deployment claim. - Fixed before scoring: no channel, feature or hyperparameter search, no score used to choose a setting, and every prediction sealed before scoring. ### Not published here Per-person, per-field, per-folder, per-clip and per-record values of any kind, including per-person minima, maxima and per-fold scores; the response-clean subset’s scores; individual permutation and shuffle replicates; confusion matrices and per-class recall; the handwriting mask-only control’s interval; identifiers, labels, features, predictions, fitted parameters and checkpoints; EEGPT weights, original or converted; measured compute; and demographics. Why cohorts this small publish so little: [small cohorts, on the data-use page →](https://bci.report/data-use/#small-cohorts-pilots) ### Credits, licences, and consent ### Orthosis error-detection EEG and EMG dataset (Kueper et al. 2024) Niklas Kueper, Kartik Chari, Judith Bütefür, Julia Habenicht, Tobias Rossol, Su Kyoung Kim, Marc Tabie, Frank Kirchner and Elsa Andrea Kirchner · EEG and EMG dataset for the detection of errors introduced by an active orthosis device (IJCAI’23 CC6 Competition), Zenodo, doi:10.5281/zenodo.8345429, CC BY 4.0. Paper: Niklas Kueper et al., EEG and EMG dataset for the detection of errors introduced by an active orthosis device, Frontiers in Human Neuroscience 18 (2024), doi:10.3389/fnhum.2024.1304311. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Ethics approval: Approved by the Ethics Committee of the Department of Computer Science and Applied Cognitive Science of the Faculty of Engineering at the University of Duisburg-Essen; no approval number is given. Informed consent: The participants gave written informed consent and were told they could stop at any time. [Dataset record ↗](https://doi.org/10.5281/zenodo.8345429) · [Paper ↗](https://doi.org/10.3389/fnhum.2024.1304311) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · [Statement read from ↗](https://europepmc.org/article/PMC/PMC10839100) ### BINGO cross-linguistic inner-speech EEG dataset (Greek and English) BINGO (Brain Imagined-Speech Communication) project, CERTH · cross-linguistic inner-speech EEG dataset, https://bingo-project.gr/datasets/; no licence is stated by the source. Documentation: Fotis P. Kalaganis, Kostas Georgiadis, Maria Kyrou, Spiros Nikolopoulos and Ioannis Kompatsiaris, D3.1 – The BINGO Benchmarking Framework, version 2.0, public deliverable (2025), https://bingo-project.gr/download/397. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Ethics approval: No ethics-approval reference is published by the source. Informed consent: The project’s public deliverable D3.1, version 3.0, states that each participant signed a consent form before the sessions. It says so for the project’s 20-participant NATO-alphabet dataset, recorded in the same CERTH laboratory; the cross-language dataset’s own documentation, D3.1 version 2.0, has no ethics or consent statement. [Dataset record ↗](https://bingo-project.gr/datasets/) · [D3.1 – The BINGO Benchmarking Framework, version 2.0 ↗](https://bingo-project.gr/download/397) · [D3.1 – The BINGO Benchmarking Framework, version 3.0 ↗](https://bingo-project.gr/download/402) · No licence stated ### EEG-fNIRS-based Handwriting Trajectory Dataset (Hugging Face) lasfk on Hugging Face · EEG-fNIRS-based Handwriting Trajectory Dataset, revision 4036680b5fb0015472d55589da2f34c5a87a3c10, CC BY-NC-SA 4.0 as declared on the dataset card, gated access. Derived analysis by BCI Report; not endorsed by the uploader. **Ethics and consent:** No ethics-approval or informed-consent statement was found. The dataset card has none; a participating team’s challenge paper, which identifies the competition as the ACM MM 2026 Multimodal Brain-Computer Interface Grand Challenge, has none; and the organisers’ paper is listed there as forthcoming. The aggregates are published with this gap stated, by the owner’s decision, to be rechecked when the organisers’ paper appears. [Dataset record ↗](https://huggingface.co/datasets/lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset) · [A participating team’s challenge paper ↗](https://arxiv.org/abs/2608.03176v2) · [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) ### REFED: real-time dynamic labelled EEG-fNIRS emotion dataset (Ning et al. 2025) Xiaojun Ning, Jing Wang, Zhiyang Feng, Tianzuo Xin, Shuo Zhang, Shaoqi Zhang, Zheng Lian, Yi Ding, Youfang Lin and Ziyu Jia · REFED: A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset, NeurIPS 2025 Datasets and Benchmarks Track. Data: REFED2025/REFED-dataset on Hugging Face, revision 0298355c2823946ac5800b2ec12a082bf2dcd01b, CC BY-NC-SA 4.0, gated. Derived analysis by BCI Report; not endorsed by the authors. **Ethics and consent, as the source states them:** Ethics approval: Approved by the local Institutional Review Board; the paper’s datasheet names the approval IRB-CASIA. Informed consent: All participants gave informed consent to the recording of their physiological signals and rating scales, and agreed that the data would be used only for non-commercial scientific research. [Dataset record ↗](https://huggingface.co/datasets/REFED2025/REFED-dataset) · [Paper ↗](https://papers.neurips.cc/paper_files/paper/2025/file/2bf0ed7c35d9d84128e7f7c72ab76402-Paper-Datasets_and_Benchmarks_Track.pdf) · [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) ### EEGPT, the model in the Orthosis comparison Paper: [EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals ↗](https://openreview.net/forum?id=XNiFuoqTuD); code: [BINE022/EEGPT ↗](https://github.com/BINE022/EEGPT) at `a0e0a8fad729e2ecf4eedb3a81548a6e6d48a705`; checkpoint `eegpt_mcae_58chs_4s_large4E.ckpt`, SHA-256 `9d63ebc80bba1ad5e3277df4eecfa05d2745e6243a12ff743c4c0a93930dfac0`. Apache-2.0 covers the repository’s source code only; the checkpoint’s own licence is unresolved, and no original or converted weights are distributed. Derived analyses by BCI Report, aggregates only; not endorsed by the data’s authors, and no raw data are redistributed. Data source: [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) · schema bci-report-pilot-studies-update-v1. Human · scalp EEG · mental arithmetic and rest · EEGMAT · NeuroLM-B, frozen ## A separate panel: NeuroLM’s frozen features on arithmetic and rest Locally executed experiment · offline · whole participants held out · not counted among the pilots above, and on no leaderboard Do NeuroLM’s frozen EEG features separate rest from mental arithmetic better than band power? NeuroLM-B stays frozen. One fixed ridge head (alpha 100) is fitted on each of three representations of the same 2,160 two-second windows from 36 people, in five participant-held-out folds: 15 fits and no tuning. Scores are means over people, each weighted equally; chance is 50.00% balanced accuracy. Equal-person means over the 36 people, each scored by heads fitted without them. Each representation has a point estimate only; the intervals belong to the paired differences below. | Representation | Dimension | Balanced accuracy | Macro-F1 | | --- | --- | --- | --- | | Tokenizer features | 768 | 64.49% | 62.52% | | Band power | 95 | 63.89% | 59.08% | | GPT hidden states | 768 | 61.20% | 59.34% | **Paired differences in balanced accuracy, in percentage points, over the same people.** Dot: the mean of the people’s paired differences; line: its 95% participant-bootstrap interval; dashed line: no difference. - Tokenizer features minus band power — the panel’s question · 95% interval −5.32 to +6.44 pp: +0.60 pp - GPT hidden states minus tokenizer features — descriptive · 95% interval −5.37 to −1.02 pp: −3.29 pp - GPT hidden states minus band power — descriptive · 95% interval −8.34 to +3.10 pp: −2.69 pp **+0.60 pp** Tokenizer features minus band power, balanced accuracy, paired over the same 36 people; 95% interval −5.32 to +6.44 ### The interval spans zero Balanced accuracy: tokenizer features 64.49%, matched log band power 63.89%, final GPT hidden states 61.20%. Macro-F1: 62.52%, 59.08% and 59.34%. Tokenizer features minus band power is +0.60 points (paired 95% participant-bootstrap interval −5.32 to +6.44): the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. GPT hidden states minus tokenizer features is −3.29 points (−5.37 to −1.02): descriptive, from this one cohort. EEGMAT is NeuroLM’s Workload task, one of the six instruction tasks in its pinned code, though not one of the 16 datasets in its published pretraining list. The training stage of this checkpoint is unresolved, so the backbone may have seen these people: pretraining exposure is unknown. This pilot has its own adapter, which filters 0.1–75 Hz and resamples to 200 Hz with no notch filter, where NeuroLM’s own pipeline adds one at 50/60 Hz; the core Arithmetic & rest matrix filters nothing and resamples to 250 Hz, so these scores have their own panel. Rest is recorded before arithmetic, so recording order moves with the label. Intervals come from 1,000 participant resamples of the 15 fixed fits and exclude refit uncertainty. [The core Arithmetic & rest matrix →](https://bci.report/protocols/arithmetic-rest/) Paired differences over the same 36 people, with 95% participant-bootstrap intervals; every contrast uses the same resamples. | Paired difference | Role | Balanced accuracy | Macro-F1 | | --- | --- | --- | --- | | Tokenizer features minus band power | the panel’s question | +0.60 pp (95% interval −5.32 to +6.44 pp · spans zero) | +3.44 pp (95% interval −3.93 to +10.67 pp · spans zero) | | GPT hidden states minus tokenizer features | descriptive | −3.29 pp (95% interval −5.37 to −1.02 pp · below zero) | −3.18 pp (95% interval −5.42 to −0.75 pp · below zero) | | GPT hidden states minus band power | descriptive | −2.69 pp (95% interval −8.34 to +3.10 pp · spans zero) | +0.26 pp (95% interval −6.64 to +7.36 pp · spans zero) | EEGMAT 1.0.0: Open Data Commons Attribution License 1.0 (Zyma et al., 2019; PhysioNet). NeuroLM code: MIT. NeuroLM-B checkpoint: CC BY 4.0 per its Hugging Face metadata; no weights distributed. [Dataset page →](https://bci.report/datasets/eegmat/) · [Credit, licence and consent ↓](https://bci.report/topics/pilot-studies/#neurolm-credits) The independent audit refitted all 15 heads from the sealed NeuroLM and band-power values and reproduced every score and interval; the NeuroLM features themselves were not re-extracted by a second implementation. ### How the panel was produced ### Windows and folds The first 60 s of each of the 72 recordings (one at rest and one during serial subtraction per person), cut into 30 two-second windows on 19 channels and resampled from 500 to 200 Hz. Labels come from the recording condition in the file name. Whole participants are held out: 8/7/7/7/7 people in the five folds. ### The head A standard scaler fitted on the training windows only, then a ridge classifier with alpha 100, balanced class weights and no hyperparameter search: 15 fits, each prediction sealed before scoring. The same head reads all three representations. ### The model NeuroLM-B, 255,107,712 parameters in a strict restore of the checkpoint, kept frozen: no fine-tuning, no text generation and no language-model head. Its outputs feed the head trained on the other folds’ labels, a supervised transfer. ### The three representations - **Tokenizer features**: the mean of NeuroLM-B’s 38 tokenizer hidden states per window (two one-second blocks of 19 channels), taken before its vocabulary projection - **Band power**: 95 absolute log band powers: 19 channels by 5 bands (1–4, 4–8, 8–13, 13–30 and 30–75 Hz), Welch with a one-second Hann window at 200 Hz on the same windows - **GPT hidden states**: the mean of the 38 final normalised GPT hidden states from the same EEG-only forward pass, which stops before the language-model head ### What else bounds this panel - The band-power arm is 95 absolute log band powers (19 channels by 5 bands, 1–75 Hz, Welch at 200 Hz), a different feature set from the core matrix’s Spectral ridge row, which uses log relative spectra from 0.5–45 Hz at 250 Hz. - The 95 band-power and 768 NeuroLM dimensions share one fixed alpha, which does not equalise capacity or effective regularisation. - There is no random-initialisation control, so the pilot says nothing about a causal benefit of pretraining. - EEGMAT is a known development cohort, used across this site. Predictions were sealed before scoring for procedural integrity: the labels are in the file names, and each person’s windows train the heads of the other folds. - The source README reports prior ICA cleaning of the published signals. - Frozen features with a supervised linear head: no text prompt, generated answer or language-model output entered any score, and nothing here measures instruction tuning, question answering, zero-shot use, cognitive workload, clinical use or online decoding. ### Not published here - Per-person, per-fold and per-window values of any kind, fold assignments, participant identifiers, condition labels, features, fitted heads, predictions and the sealed prediction files: only equal-person means, paired differences with their intervals and cohort counts are published. - Demographics from the source dataset’s participant file. - Measured compute: worker times, memory and process counts from the handoff, the receipts and the audits. [what this release publishes, on the data-use page →](https://bci.report/data-use/#sources-neurolm-eegmat) ### Credits, licences and consent ### EEGMAT, the recordings Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. **Ethics and consent, as the source states them:** Paper reports Bioethics Commission approval (15 August 2018) and written consent from every participant. **Use:** Personal noncommercial research; aggregate results only [Dataset record ↗](https://physionet.org/content/eegmat/1.0.0/) · [Open Data Commons Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/) · [Dataset page →](https://bci.report/datasets/eegmat/) ### NeuroLM, the model Paper: [NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals ↗](https://arxiv.org/abs/2409.00101); Code: [935963004/NeuroLM ↗](https://github.com/935963004/NeuroLM/tree/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177) at `0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177`, [MIT](https://github.com/935963004/NeuroLM/blob/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177/LICENSE). Checkpoint: [`checkpoints/NeuroLM-B.pt` ↗](https://huggingface.co/Weibang/NeuroLM/tree/eddfff5c64a4139442f826d6c67c8369fd00f45a) at `eddfff5c64a4139442f826d6c67c8369fd00f45a`, SHA-256 `ffe098bc138b89f8817d3710a3604498d8ecd15135080e2ca27735d05c6d29ab`. CC BY 4.0, as the licence tag of its Hugging Face repository states; its 30-byte model card adds no terms. No weights are distributed; the SHA-256 identifies the checkpoint used and grants nothing. Derived analysis by BCI Report, aggregates only; not endorsed by NeuroLM’s authors or the data’s authors, and no raw data are redistributed. Data source: [neurolm-eegmat-update.json](https://bci.report/data/neurolm-eegmat-update.json) · schema bci-report-neurolm-eegmat-update-v1. Human · scalp EEG · arithmetic and rest, motor imagery, P300 · NeuroRVQ, frozen ## A separate panel: NeuroRVQ’s frozen backbone on three tasks Locally executed experiment · offline · whole participants held out · outside the pilot count above Does NeuroRVQ’s frozen backbone, read by a fitted ridge head, separate the classes of three existing tasks better than simple features of the same windows? The backbone stays frozen. On each task one ridge head (alpha 100) is fitted per fold on the other participants, for the backbone’s 800 values and for simple features of the same windows: 36 people on EEGMAT, 10 on motor imagery against rest and 21 on P300 target against non-target, in 5 participant-held-out folds, nothing tuned. Scores are participant means; chance is 50.00% balanced accuracy. Each task has one paired difference with a participant-bootstrap interval. **Paired differences in balanced accuracy, the NeuroRVQ backbone minus each task’s primary comparator, in percentage points.** Dot: the mean of the people’s paired differences; line: its 95% participant-bootstrap interval; dashed line: no difference. Each task has its own people, windows and comparator. - Arithmetic and rest · EEGMAT — minus band power · 95% interval −3.52 to +6.67 pp: +1.53 pp - Motor imagery and rest · ds003810 — minus temporal means · 95% interval −20.17 to −10.25 pp: −15.25 pp - P300 target and non-target · ds006593 — minus temporal means · 95% interval −5.56 to −1.51 pp: −3.57 pp ### Arithmetic and rest · EEGMAT 36 people, 72 recordings, 2,160 two-second windows (30 from the first 60 s of each recording), 19 channels, 5 folds of whole participants; its own input chain, resampled from 500 to 200 Hz. [This task’s core matrix →](https://bci.report/protocols/arithmetic-rest/) Participant means over the 36 people, each scored by heads fitted without them. | Representation | Role | Dimension | Balanced accuracy | Macro-F1 | | --- | --- | --- | --- | --- | | NeuroRVQ backbone | probe | 800 | 66.34% | 64.82% | | Band power | matched comparator | 95 | 64.81% | 60.50% | **+1.53 pp** NeuroRVQ backbone minus band power, balanced accuracy, paired over the same 36 people ### The interval spans zero NeuroRVQ backbone minus band power: +1.53 points, paired 95% participant-bootstrap interval −3.52 to +6.67; the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. - **NeuroRVQ backbone**: the four 200-value branch outputs of the frozen backbone, each averaged over the window’s 38 tokens (19 channels by two one-second patches) and joined into 800 values - **Band power**: 95 log band powers: 19 channels by five bands (1–4, 4–8, 8–13, 13–30 and 30–44.5 Hz), Welch with one-second Hann windows at 200 Hz on the same adapted windows - EEGMAT has its own input chain: the first 60 seconds of each recording, notches at 50, 60 and 100 Hz, a third-order 0.5–44.5 Hz band-pass run forward and backward, clipping at ±500 µV and resampling from 500 to 200 Hz, with no demeaning or re-referencing. The band-power arm reads the same adapted windows. - EEGMAT’s recording order (rest before arithmetic), the source’s prior ICA cleaning and EEGMAT’s use as a development cohort are [stated in the NeuroLM panel above ↑](https://bci.report/topics/pilot-studies/#neurolm-eegmat) and bound this block too: it is not an untouched external confirmation. The independent audit rebuilt every row, fold and label, refitted all 10 heads and reproduced all 4,320 held-out predictions, the scores and the interval; it read both arms’ values from the producer’s sealed files and did not re-extract them. ### Motor imagery and rest · ds003810 10 people, 1,200 two-second windows, 15 channels, 5 folds of whole participants: the core matrix’s prepared windows and folds, resampled from 125 to 200 Hz, with the comparators recomputed in that chain. [This task’s core matrix →](https://bci.report/protocols/mi-rest/) Participant means over the 10 people, each scored by heads fitted without them. | Representation | Role | Dimension | Balanced accuracy | Macro-F1 | | --- | --- | --- | --- | --- | | NeuroRVQ backbone | probe | 800 | 55.08% | 52.64% | | Temporal means | primary comparator | 1,500 | 70.33% | 69.84% | | Band power | descriptive comparator | 75 | 58.83% | 56.48% | **−15.25 pp** NeuroRVQ backbone minus temporal means, balanced accuracy, paired over the same 10 people ### The interval lies below zero NeuroRVQ backbone minus temporal means: −15.25 points, paired 95% participant-bootstrap interval −20.17 to −10.25; the interval lies below zero; descriptive, from this one cohort. - **NeuroRVQ backbone**: the four 200-value branch outputs of the frozen backbone, each averaged over the window’s 30 tokens (15 channels by two one-second patches) and joined into 800 values - **Temporal means**: 1,500 means of non-overlapping four-sample (20 ms) bins after resampling to 200 Hz: 15 channels by 100 bins - **Band power**: 75 log band powers: 15 channels by five bands (1–4, 4–8, 8–13, 13–30 and 30–45 Hz), Welch with one-second windows at 200 Hz - Motor imagery and P300 reuse the core matrix’s prepared windows, people and fold rule, resampled per window from 125 and 300 Hz to 200 Hz with no further filter, clipping or rescaling; motor imagery keeps the core per-channel epoch demeaning. Their temporal and band-power arms were recomputed in this chain, apart from the matrix run: the motor-imagery band power differs from that protocol’s Spectral ridge row, and the P300 temporal means from that protocol’s Temporal ridge row. - Motor-imagery and P300 amplitudes are read as microvolts from the source cache: ds003810’s EDF unit fields are blank and its channel files declare microvolts, and ds006593’s headers give a 0.1 µV resolution with blank reference fields; the P300 cache’s original preparation from the raw files was not re-run. Neither reading is an independent instrument calibration. The independent audit recomputed the temporal and band-power values bit for bit, refitted all 15 heads and reproduced all 3,600 held-out predictions, the scores and the interval; a supplementary check re-extracted all 1,200 × 800 NeuroRVQ values exactly, reusing the accepted backbone loader. ### P300 target and non-target · ds006593 21 people, 2,520 one-second epochs (60 target and 60 non-target per person), 19 channels, 5 folds of whole participants: the core matrix’s prepared windows and folds, resampled from 300 to 200 Hz, with the comparators recomputed in that chain. [This task’s core matrix →](https://bci.report/protocols/p300-target/) Participant means over the 21 people, each scored by heads fitted without them. | Representation | Role | Dimension | Balanced accuracy | Macro-F1 | | --- | --- | --- | --- | --- | | NeuroRVQ backbone | probe | 800 | 50.40% | 47.36% | | Temporal means | primary comparator | 950 | 53.97% | 52.21% | **−3.57 pp** NeuroRVQ backbone minus temporal means, balanced accuracy, paired over the same 21 people ### The interval lies below zero NeuroRVQ backbone minus temporal means: −3.57 points, paired 95% participant-bootstrap interval −5.56 to −1.51; the interval lies below zero; descriptive, from this one cohort. - **NeuroRVQ backbone**: the four 200-value branch outputs of the frozen backbone, each averaged over the window’s 19 tokens (one one-second patch per channel) and joined into 800 values - **Temporal means**: 950 means of non-overlapping four-sample (20 ms) bins after resampling to 200 Hz: 19 channels by 50 bins - The P300 sample is balanced at 60 target and 60 non-target epochs per person, where the source inquiries hold about one target in ten letters, and stimuli come about every 0.3 s, so each one-second window holds later responses. The scores are for single epochs and do not measure speller accuracy. A separately written audit re-extracted all 2,520 × 800 NeuroRVQ values and 2,520 × 950 temporal means exactly, reusing the accepted backbone loader, refitted all 10 heads and reproduced all 5,040 held-out predictions, the scores and the interval. ### The model and its head - Every score uses the released NeuroRVQ backbone with parameter-free branch normalisation and mean-token pooling: 263 trained tensors restore strictly, the four branch normalisations carry no parameters and take the source’s default scale 1 and shift 0, and each branch’s tokens are averaged and the four 200-value branches joined into 800 values. The authors’ inference wrapper fails a strict restore, because eight normalisation tensors it expects are missing from the checkpoint. - Supervised transfer: the backbone stays frozen, and in each fold a standardiser and a ridge head (alpha 100, balanced class weights) are fitted on the other participants’ windows. No score is zero-shot, and nothing was tuned. - Pretraining exposure is unknown. None of the three datasets is named in the paper’s list of 13 pretraining corpora, but the list includes a proprietary collection and no per-checkpoint list of corpora or participants is published, so the backbone may have seen these recordings. The head: a standard scaler fitted on the training participants’ windows only, then a ridge classifier with alpha 100, the Cholesky solver, balanced class weights and no hyperparameter search; 10, 15 and 10 fits on the three tasks, each prediction sealed before scoring. ### What else bounds this panel - The checkpoint and the paper do not establish the amplitude scale or reference NeuroRVQ was pretrained on. Every chain here feeds microvolts as an adaptation: the EEGMAT chain follows the notch, band-pass and clip of the example preprocessing in NeuroRVQ’s code, and the motor-imagery and P300 chains leave all three out. No score here reproduces NeuroRVQ’s published pipeline, and none separates the representation from an input mismatch. - The arms differ in size under one fixed alpha: NeuroRVQ’s 800 values against 95 band powers on EEGMAT, 1,500 temporal means and 75 band powers on motor imagery, and 950 temporal means on P300. There is no random-initialisation arm, so the panel says nothing about what pretraining contributes. - Each paired difference carries a descriptive 95% interval from 20,000 paired participant resamples, conditional on the fixed folds and fitted heads; each arm has a point estimate only. - Each panel on this page has its own input chain and band-power arm, so these scores stay outside every leaderboard and are not ranked against the NeuroLM panel, the ten-second panel or the core matrix. - Offline results for one frozen variant: the EEGMAT chain filters the retained first minute of each recording as one block before cutting it into windows, and nothing here measures cognitive workload, clinical use, online decoding or other uses of NeuroRVQ. Protocols not run, with what each still needs: Semantic target ERP (the original calibration evidence is still missing); SSVEP · 4 channels (physical units and reference are unresolved); SSVEP · 8 channels (physical units are unresolved, and PO5 and PO6 are missing from the model’s channel list); Idle & command (its inputs are missing, and it needs its own calibration, threshold and split protocol); Sleep staging (deferred to a separate study). ### Not published here - Per-person, per-fold and per-window values of any kind, fold assignments, participant identifiers, labels, extracted values, fitted heads, predictions and the sealed files: only participant-mean scores, one paired difference per task with its interval, and cohort counts are published. - How many people each arm scored above 50%, which the sources give as descriptive point-estimate counts. [what this release publishes, on the data-use page →](https://bci.report/data-use/#sources-neurorvq) ### Credits, licences and consent ### NeuroRVQ Paper: [arXiv:2510.13068 ↗](https://arxiv.org/abs/2510.13068); code: [KonstantinosBarmpas/NeuroRVQ ↗](https://github.com/KonstantinosBarmpas/NeuroRVQ/tree/926e770d9d16b6aa308404280fa0cc0211a6f9fb) at `926e770d9d16b6aa308404280fa0cc0211a6f9fb`, [CC BY-NC 4.0](https://github.com/KonstantinosBarmpas/NeuroRVQ/blob/926e770d9d16b6aa308404280fa0cc0211a6f9fb/LICENSE). Checkpoint: [`pretrained_models/foundation_models/NeuroRVQ_EEG_foundation_model_v1.pt` ↗](https://huggingface.co/ntinosbarmpas/NeuroRVQ/tree/d944b87f44ae0ba2923b2f10d0518f23f6803b76) at `d944b87f44ae0ba2923b2f10d0518f23f6803b76`, SHA-256 `d14ef4eae2d4594d5f1f709226a3bb5f362afd84b68050d5da0fa0586aea886a`, [CC BY-NC 4.0](https://huggingface.co/ntinosbarmpas/NeuroRVQ/blob/d944b87f44ae0ba2923b2f10d0518f23f6803b76/LICENSE). NeuroRVQ’s code and weights are licensed CC BY-NC 4.0, the licence file of both repositories at the pinned revisions. Only aggregate figures are published here; no weights are distributed, and no endorsement by NeuroRVQ’s authors is implied. ### EEGMAT [Credit, ethics and consent: in the NeuroLM panel’s credits above ↑](https://bci.report/topics/pilot-studies/#neurolm-credits) · [Open Data Commons Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/) · [Dataset page →](https://bci.report/datasets/eegmat/) ### ds003810 Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ **Ethics and consent, as the source states them:** Study paper reports ethics approval by CEySTE CCT-CONICET Santa Fe, Declaration of Helsinki compliance, and signed consent. **Use:** Personal noncommercial research; aggregate results only [Dataset record ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset page →](https://bci.report/datasets/ds003810/) ### ds006593 OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. **Ethics and consent, as the source states them:** Dataset metadata/README report IRB approval #10-05-2022 and written informed consent. **Use:** Personal noncommercial research; aggregate results only [Dataset record ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset page →](https://bci.report/datasets/ds006593/) Data source: [neurorvq-update.json](https://bci.report/data/neurorvq-update.json) · schema bci-report-neurorvq-update-v1. Human · scalp EEG · mental arithmetic and rest · EEGMAT · EEG-Captioner and AXON, frozen ## A separate panel: two frozen encoders on ten-second arithmetic and rest windows Locally executed experiment · offline · whole participants held out · outside the pilot count above Do frozen EEG-Captioner or AXON features separate rest from mental arithmetic in ten-second windows better than Welch band power? Both encoders stay frozen. Each representation is read by one fixed ridge head (alpha 100) in 5 subject-disjoint folds on the same 432 ten-second windows from 36 people, and each encoder is set against Welch band power on the same windows. Scores are participant-mean balanced accuracies; each paired difference carries a participant-bootstrap interval. ### The cohort and the shared steps 36 people, 72 recordings, 432 ten-second windows, 19 channels at 200 Hz, 5 folds. Each complete recording was filtered at 0.1–75 Hz with a 50 Hz notch and resampled to 200 Hz, then its first 60 s were cut into six ten-second windows. The head: a standard scaler fitted on the training folds only, then a ridge classifier with alpha 100, the SVD solver, balanced class weights and no hyperparameter search, 5 fits per representation, each prediction sealed before scoring; the same head reads both encoders. 20,000 paired participant draws of the 36 participant differences in each subsection, with a 2.5–97.5 percentile interval. Both intervals span zero. Each interval resamples the participant differences of this fixed cohort and its fixed predictions; training sets overlap across folds, so it describes this cohort, not a population. **Source annotation QC:** EEG-Captioner’s published annotation-QC rules, run unchanged at the pinned source revision on each whole preprocessed recording, before the zero Oz channel was inserted. A recording is flagged when more than 0.3 of its channels are flagged or the summed duration of BAD annotations exceeds 0.2 of the recording. A channel is flagged by MNE’s adjacent-sample amplitude criteria (500 µV peak, 1 µV flat, 5% of the record, segments of at least 0.02 s); the 100 µV segment rule applies only inside a recording that passes. The flag is set per recording at the parent-trial level. It does not say that each window has an artifact, and it is not evidence that EEGMAT as a whole is unusable. ### EEG-Captioner encoder, mean-token 256-d + ridge - **Input:** the 19 measured channels in microvolts with a zero Oz channel inserted, as the encoder’s loader permits, divided by 100. - **Representation:** the mean of the 200 output tokens of EEG-Captioner’s frozen encoder, 256 dimensions. - **Baseline:** 95 absolute log band powers: 19 channels by 5 bands (0.5–4, 4–8, 8–13, 13–30 and 30–45 Hz), Welch at 200 Hz with 400-sample Hann segments and 200-sample overlap, on the same windows. - Only EEG-Captioner’s encoder was run; its language model, projector and language adapter were not loaded, and there is no caption result. Participant means over the 36 people, each scored by heads fitted without them. | Representation | Role | Dimension | Balanced accuracy | | --- | --- | --- | --- | | EEG-Captioner encoder, mean-token 256-d + ridge | encoder | 256 | 65.74% | | Welch band power (95-d, 0.5–45 Hz, ten-second windows) | baseline — matched for this run | 95 | 67.59% | **−1.85 pp** EEG-Captioner encoder minus Welch band power, balanced accuracy, paired over the same 36 people ### The interval spans zero EEG-Captioner encoder, mean-token 256-d + ridge: 65.74%; Welch band power (95-d, 0.5–45 Hz, ten-second windows): 67.59%. The paired difference is −1.85 points (95% participant-bootstrap interval −9.03 to +4.86): the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. 0 of 432 windows meet EEG-Captioner’s source annotation-QC rules, and 72 of 72 recordings are flagged by them. A separately operated reconstruction, importing none of the producer’s scientific helpers, reproduced all 432 encoder and baseline feature rows exactly from the sealed preprocessed arrays. It refitted every head, reproduced every hard prediction exactly with decision scores within 1e-6 of the producer’s, and matched every score and the bootstrap. The raw preprocessing and the source QC were checked against their pinned source and hashes, not recomputed. ### AXON, 512-d + ridge - **Input:** the same 19 measured channels in microvolts, reordered to AXON’s channel order and standardised per channel and window inside AXON. - **Representation:** the mean over AXON’s channel and patch tokens, 512 dimensions. Participant means over the 36 people, each scored by heads fitted without them. | Representation | Role | Dimension | Balanced accuracy | | --- | --- | --- | --- | | AXON, 512-d + ridge | encoder | 512 | 70.14% | | Welch band power (95-d, 0.5–45 Hz, ten-second windows) | baseline — the same predictions reused without refitting; not matched for normalisation or parameter count | 95 | 67.59% | **+2.55 pp** AXON minus Welch band power, balanced accuracy, paired over the same 36 people ### The interval spans zero AXON, 512-d + ridge: 70.14%; Welch band power (95-d, 0.5–45 Hz, ten-second windows): 67.59%. The paired difference is +2.55 points (95% participant-bootstrap interval −5.79 to +10.42): the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. 0 of 432 windows meet EEG-Captioner’s source annotation-QC rules, and 72 of 72 recordings are flagged by them. AXON’s own numerical input check passed 432 of 432, a different check. **The check:** every sample finite, and every channel’s sample standard deviation in every window above 1e-12 µV: numerical non-degeneracy, a separate matter from signal quality. - The Welch comparison cannot isolate what AXON’s pretraining contributes. - AXON’s interval is descriptive for a second reason too: the EEG-Captioner and Welch outcomes were known before AXON ran. - AXON reads nominal template electrode positions, not positions digitised for each person, and the cache’s 0.1–75 Hz band with a 50 Hz notch differs from the pretraining preprocessing on AXON’s model card (0.5–99.5 Hz, a 50 or 60 Hz notch, standardisation and clipping at 15 standard deviations). AXON’s first independent audit failed its feature tolerance and is kept. The corrected audit changed only the auditor’s frozen-parameter state to match production; it reproduced all 432 production rows byte for byte under the unchanged tolerance, refitted the heads and matched every score. Production was unchanged. No EEG-Captioner against AXON difference is computed: AXON ran after the EEG-Captioner and Welch outcomes were known, and the two encoders read different inputs (a zero Oz channel and division by 100 for EEG-Captioner; reordered channels with its own standardisation for AXON). Not ranked against the two-second results on this site: the core Arithmetic & rest matrix, the REVE and EEGPT frozen rows, LaBraM last-block adaptation, the NeuroLM panel and the NeuroRVQ panel. These probes use filtered ten-second windows at 200 Hz, where the core protocol uses unfiltered two-second windows at 250 Hz, and its Spectral ridge row uses log relative spectra from 0.5–45 Hz rather than these absolute band powers. [The core Arithmetic & rest matrix →](https://bci.report/protocols/arithmetic-rest/) ### What else bounds this panel - EEGMAT’s recording order (rest before arithmetic), [stated in the NeuroLM panel above ↑](https://bci.report/topics/pilot-studies/#neurolm-eegmat), holds for these windows too: time and artifacts also move with the label, subject-disjoint folds do not remove this, and neither probe isolates cognitive workload. - Whether either encoder saw EEGMAT in pretraining is unknown, so exposure is shown as unknown for both. - EEGMAT’s linked-ear reference and its upstream filtering are uncertain from the source files, and further filtering cannot restore frequencies removed before release. - Frozen encoders with a supervised linear head, scored on recording condition: the filters ran over whole recordings, using context outside each window, and nothing here measures clinical use or online decoding. ### Not published here - Per-person, per-fold and per-window values of any kind: per-fold balanced accuracy, fold assignments, participant identifiers, features, fitted heads, predictions and the sealed files. Only participant-mean balanced accuracy, paired differences with their intervals, cohort counts and the source-QC totals are published. [what this release publishes, on the data-use page →](https://bci.report/data-use/#sources-eegmat-ten-second) ### Credits, licences and consent ### EEGMAT, the recordings [Credit, ethics and consent: in the NeuroLM panel’s credits above ↑](https://bci.report/topics/pilot-studies/#neurolm-credits) · [Open Data Commons Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/) · [Dataset page →](https://bci.report/datasets/eegmat/) ### EEG-Captioner, the encoder Model: [LyJinAoki/EEG-Captioner ↗](https://huggingface.co/LyJinAoki/EEG-Captioner) at `1d56a371d68a108dd56c362cc41205c5b511cc76`; code: [lyJinAoki/EEG-Captioner ↗](https://github.com/lyJinAoki/EEG-Captioner) at `088d9af2dedcb181534f7fdcff7c913a19e10f04`. Encoder: SHA-256 `060a5c0d9ac11543470038adbcbc5d9ba5828c1dc7c7d5097549435fbe28e989`. EEG-Captioner’s code: MIT. EEG-Captioner’s encoder weights: Apache-2.0, as stated by the portable weight metadata of the Hugging Face repository at the pinned revision. ### AXON Model: [MannasAI/axon-eeg ↗](https://huggingface.co/MannasAI/axon-eeg) at `fe7f85239e6018f393039a10a103bf7075c93aa6`; checkpoint `model.safetensors`, SHA-256 `89c864a3ed947588ffc740cd4b342a2158661b13898c0dd9a86abebd33ffc882`. Code: [braindecode/braindecode ↗](https://github.com/braindecode/braindecode) at `8c618f15a702f63d4899e3e8bc2e6f10a5f54151`, module `braindecode.models.AXON`. AXON’s checkpoint: Apache-2.0, as stated by the licence tag of the Hugging Face repository at the pinned revision. The AXON module: Apache-2.0, the per-file terms of the AXON module in Braindecode, within a BSD-3-Clause project (the licence at the Braindecode project root). No weights are distributed; each SHA-256 identifies the file used and grants nothing. Derived analysis by BCI Report, aggregates only; not endorsed by EEG-Captioner’s or AXON’s authors or the data’s authors, and no raw data are redistributed. Data source: [eegmat-ten-second-update.json](https://bci.report/data/eegmat-ten-second-update.json) · schema bci-report-eegmat-ten-second-update-v1. Human · scalp EEG · mental arithmetic and rest; motor imagery and rest · EEGMAT, ds003810 · a local model and two frozen encoders ## A separate panel: the components of Neural State Prediction, tested locally Locally executed experiment · offline · whole participants held out · a panel of its own after the counted pilots Do the target components Neural State Prediction proposes change decoding in a small model trained here, and how much of a frozen encoder’s drop under zeroed channels returns when its head is refitted? Neural State Prediction (NSP), an EEG pretraining proposal cited below, builds its prediction target with channel and time position subtracted (ResID) and keeps the context apart from the target’s neighbourhood (topology). Block 1 trains a model of 1,378,105 parameters on EEGMAT with each component on or off; block 2 sets the trained models against the same model never trained; block 3 zeroes channels at the input of two published eeg-fm-masking encoders and of a log band-power ridge. Every score is an equal-person mean on people held out whole; each difference carries a descriptive paired-participant interval. ### Block 1 · The components in a small local model 1,378,105 parameters (4 layers, width 128), trained on EEGMAT alone: 36 people, 2,160 two-second windows, 19 channels resampled to 200 Hz. 4 cells × 5 participant-disjoint folds × 3 paired seeds = 60 jobs of 500 steps. Each fold’s test people are left out of the unsupervised training and of the ridge readout (alpha 100, read at the final step). Balanced accuracy, equal-person means over the 36 people, each scored by heads fitted without them. | Cell | Target residualised (ResID) | Context geometry | Balanced accuracy | | --- | --- | --- | --- | | ResID0_Random | ResID off | random | 58.72% | | ResID1_Random | ResID on | random | 58.44% | | ResID0_Topology | ResID off | topology | 59.12% | | ResID1_Topology | ResID on | topology | 58.78% | **+0.34 pp** with ResID on, topology minus random ### Each interval spans zero Fixed in advance as the primary contrast, with ResID on, topology minus random: +0.34 pp (descriptive 95% interval −0.17 to +0.86). With topology, ResID on minus off: −0.34 pp (−1.23 to +0.60); the interaction: −0.06 pp (−0.71 to +0.69). Each interval spans zero and no point estimate reaches the 2-point convention fixed in advance: no clear downstream gain was established under this local, fixed protocol. Linear probes for channel and time position, mean of the 15 jobs in each cell; chance is 5.26% for channel and 10.0% for time. | Cell | Target · channel | Context · channel | Target · time | Context · time | | --- | --- | --- | --- | --- | | ResID0_Random | 76.7% | 65.9% | 10.6% | 10.4% | | ResID1_Random | 8.4% | 62.4% | 10.3% | 10.2% | | ResID0_Topology | 77.1% | 66.7% | 10.5% | 10.4% | | ResID1_Topology | 8.4% | 62.5% | 10.3% | 10.2% | ResID moved the probe of the Target’s channel from 76.7% and 77.1% to 8.4% and 8.4%, still above chance; the Context’s channel probe stayed between 62.4% and 66.7%; the time probes stayed between 10.2% and 10.6% in every cell. The Target’s position statistics changed sharply. Drop in mean cosine similarity between each model’s predictions and its own fixed Target when its input is changed. Held-out test people only: within each seed the folds are weighted by their test people, then the paired seeds are averaged. | Cell | Input zeroed | Another person’s input | | --- | --- | --- | | ResID0_Random | 0.355 | 0.385 | | ResID1_Random | 0.572 | 0.586 | | ResID0_Topology | 0.306 | 0.317 | | ResID1_Topology | 0.502 | 0.514 | A fixed-Target intervention supports dependence on visible content. ### Limits of this block - A local model of 1,378,105 parameters, trained on one cohort for 500 steps per job: not the authors’ model, weights or setting (about 42.83M parameters pretrained on TUEG and evaluated on 30 downstream datasets), so it neither replicates nor refutes NSP. - Intervals are descriptive paired-participant bootstraps (10,000 draws) over overlapping training folds and 3 seeds, with no p-value; the 60 jobs are not 60 independent samples. - The 2-point line is a convention fixed before the runs, not an equivalence bound. - Topology follows the reference code’s rule (a spatial expansion, then a temporal one, cross terms included), not the paper’s simpler union formula. - The position probes read channel and time position, not participant identity, and a lower probe of the Target’s channel does not by itself establish a purer neural state. - Each drop compares one model with its own fixed Target, so it is read within a cell, not across cells; the absolute cosines are not comparable between cells and are not published. The independent numeric audit replayed 60 task heads, 300 diagnostic heads, 25,920 predictions and 30,000 training steps; it did not recompute encoder outputs or optimizer updates. The first audit of this block stopped as incomplete because the auditor’s own reference rounded a training-only statistic in single precision; it is kept. A corrected auditor, checked independently, then passed all 60 jobs in double precision with the same tolerances; the producer and the aggregate are unchanged. ### Block 2 · The trained model against the same model never trained The same architecture from the same initial states as the four trained cells, never trained, read by the same ridge head. Each head is fitted on 10%, 25% or 100% of the outer-training windows; analysed after the main results were known. Balanced accuracy by the share of outer-training windows that fit the head; equal-person means over the 36 people. | Model | 10% of labels | 25% of labels | 100% of labels | | --- | --- | --- | --- | | ResID0_Random | 55.25% | 56.20% | 58.72% | | ResID1_Random | 55.11% | 55.62% | 58.44% | | ResID0_Topology | 55.20% | 56.22% | 59.12% | | ResID1_Topology | 55.37% | 55.59% | 58.78% | | Never trained | 53.78% | 55.17% | 56.62% | Trained minus never trained, with descriptive 95% intervals. | Model | 10% of labels | 25% of labels | 100% of labels | | --- | --- | --- | --- | | ResID0_Random | +1.47 pp (+0.31 to +2.59) | +1.03 pp (−0.28 to +2.45) | +2.10 pp (+0.37 to +3.94) | | ResID1_Random | +1.33 pp (+0.09 to +2.52) | +0.45 pp (−0.85 to +1.84) | +1.82 pp (−0.08 to +3.67) | | ResID0_Topology | +1.42 pp (+0.20 to +2.61) | +1.05 pp (−0.26 to +2.41) | +2.50 pp (+0.80 to +4.24) | | ResID1_Topology | +1.59 pp (+0.37 to +2.82) | +0.42 pp (−0.94 to +1.85) | +2.16 pp (+0.19 to +4.15) | ### Positive on the mean at every budget All twelve trained-minus-untrained differences are positive on the mean. With 10% of the labels, all four intervals lie above zero; with 25%, all four span zero; with every label the differences run from +1.82 pp to +2.50 pp, and the interval of ResID1_Random (−0.08 to +3.67) spans zero. The three component contrasts at each fitting-label budget: 10% and 25% analysed after the main results, every label fixed in advance (block 1). | Contrast | 10% of labels | 25% of labels | Every label (block 1) | | --- | --- | --- | --- | | With ResID on: topology minus random | +0.26 pp (−0.19 to +0.73) | −0.03 pp (−0.43 to +0.37) | +0.34 pp (−0.17 to +0.86) | | With topology: ResID on minus off | +0.17 pp (−0.74 to +1.06) | −0.63 pp (−1.53 to +0.29) | −0.34 pp (−1.23 to +0.60) | | ResID × geometry interaction | +0.31 pp (−0.32 to +0.94) | −0.05 pp (−0.66 to +0.57) | −0.06 pp (−0.71 to +0.69) | All six component contrasts with heads fitted on part of the labels span zero: no clear component gain was established at any measured fitting-label budget. ### Limits of this block - These 18 comparisons were made after the main results; their intervals are descriptive, conditional on fixed label subsets, overlapping folds and three seeds. - The baseline is the untrained model at 56.62% with every label, not the 50% chance level. - A fitting-label budget is the share of outer-training windows that fit the scaler and the head; every training person is kept and the full labels set the stratification, so it is not a label-acquisition cost. - An independent forward replay of 4 final checkpoints, one per cell, matched all 40 array comparisons with a maximum scaled error of 0; it covers 4 of the 60 checkpoints and no optimizer step, and read no task label. The numeric audit refitted 165 new heads, re-checked the 60 reused ones and replayed 97,200 predictions. ### Block 3 · Channels zeroed on two published encoders, head kept or refitted The frozen eeg-fm-masking encoders MAE and JEPA (r = 9 cm, L = 2), and a log band-power ridge (5 bands, visible channels): five log band powers per visible channel (1–4, 4–8, 8–13, 13–30 and 30–45 Hz, from a Hann periodogram of each two-second window at 200 Hz) with an indicator for each zeroed channel, read by the same ridge head (alpha 100). On EEGMAT (36 people, 6 of 19 channels zeroed in each window) and on ds003810’s motor imagery against rest (10 people, 5 of 15 zeroed). Each window gets 8 label-independent draws of its nearest channels (near) and of as many random channels (random), shared by the three models. - **Head kept:** The scaler and ridge head trained on clean windows, kept fixed for the zeroed windows. - **Head refitted:** A scaler and ridge head refitted on the zeroed training windows of the outer-training people only, each original window’s 8 views weighted equally. Balanced accuracy, equal-person means; chance is 50%. | Task · model | All channels | Near zeroed, head kept | Near zeroed, head refitted | Random zeroed, head kept | Random zeroed, head refitted | | --- | --- | --- | --- | --- | --- | | EEGMAT · MAE | 63.56% | 54.31% | 64.31% | 55.03% | 63.99% | | EEGMAT · JEPA | 61.53% | 54.17% | 63.96% | 52.72% | 64.19% | | EEGMAT · Log band-power ridge | 64.95% | 53.39% | 64.21% | 52.88% | 65.05% | | ds003810 · MAE | 68.08% | 56.55% | 67.38% | 57.54% | 65.93% | | ds003810 · JEPA | 63.17% | 53.35% | 61.00% | 53.95% | 59.35% | | ds003810 · Log band-power ridge | 57.33% | 52.67% | 55.68% | 51.89% | 55.29% | ### Two prespecified contrasts Fixed in advance on EEGMAT with the clean-trained head, near minus random zeroing: MAE −0.72 pp (−1.71 to +0.30); JEPA +1.45 pp (+0.17 to +2.71). With the head refitted, EEGMAT’s near-zeroed scores return to 64.31% for MAE and 63.96% for JEPA, against 63.56% and 61.53% on all channels; the band-power ridge’s near-zeroed score moves from 53.39% to 64.21%. The other thirty-four contrasts The other thirty-four contrasts | Task | Model | Head | Contrast | Difference (95% interval) | | --- | --- | --- | --- | --- | | EEGMAT | MAE | kept | near zeroed minus all channels | −9.25 pp (−13.07 to −5.34) | | EEGMAT | MAE | kept | random zeroed minus all channels | −8.54 pp (−12.09 to −5.05) | | EEGMAT | MAE | refitted | near minus random zeroing | +0.32 pp (−1.23 to +1.90) | | EEGMAT | MAE | refitted | near zeroed minus all channels | +0.75 pp (−1.20 to +2.71) | | EEGMAT | MAE | refitted | random zeroed minus all channels | +0.42 pp (−1.97 to +3.02) | | EEGMAT | JEPA | kept | near zeroed minus all channels | −7.36 pp (−11.24 to −3.08) | | EEGMAT | JEPA | kept | random zeroed minus all channels | −8.81 pp (−12.94 to −4.24) | | EEGMAT | JEPA | refitted | near minus random zeroing | −0.23 pp (−1.82 to +1.30) | | EEGMAT | JEPA | refitted | near zeroed minus all channels | +2.43 pp (+0.24 to +4.80) | | EEGMAT | JEPA | refitted | random zeroed minus all channels | +2.66 pp (−0.31 to +5.87) | | EEGMAT | Log band-power ridge | kept | near minus random zeroing | +0.51 pp (−0.75 to +1.78) | | EEGMAT | Log band-power ridge | kept | near zeroed minus all channels | −11.56 pp (−15.60 to −7.29) | | EEGMAT | Log band-power ridge | kept | random zeroed minus all channels | −12.07 pp (−16.04 to −7.98) | | EEGMAT | Log band-power ridge | refitted | near minus random zeroing | −0.84 pp (−2.22 to +0.42) | | EEGMAT | Log band-power ridge | refitted | near zeroed minus all channels | −0.74 pp (−3.43 to +2.04) | | EEGMAT | Log band-power ridge | refitted | random zeroed minus all channels | +0.10 pp (−2.90 to +3.26) | | ds003810 | MAE | kept | near minus random zeroing | −0.99 pp (−2.33 to +0.32) | | ds003810 | MAE | kept | near zeroed minus all channels | −11.53 pp (−14.40 to −8.66) | | ds003810 | MAE | kept | random zeroed minus all channels | −10.54 pp (−12.67 to −8.49) | | ds003810 | MAE | refitted | near minus random zeroing | +1.45 pp (+0.24 to +2.70) | | ds003810 | MAE | refitted | near zeroed minus all channels | −0.71 pp (−2.13 to +0.39) | | ds003810 | MAE | refitted | random zeroed minus all channels | −2.16 pp (−4.01 to −0.40) | | ds003810 | JEPA | kept | near minus random zeroing | −0.59 pp (−1.77 to +0.54) | | ds003810 | JEPA | kept | near zeroed minus all channels | −9.81 pp (−14.07 to −5.41) | | ds003810 | JEPA | kept | random zeroed minus all channels | −9.22 pp (−12.79 to −5.57) | | ds003810 | JEPA | refitted | near minus random zeroing | +1.65 pp (−0.68 to +4.00) | | ds003810 | JEPA | refitted | near zeroed minus all channels | −2.17 pp (−4.69 to +0.63) | | ds003810 | JEPA | refitted | random zeroed minus all channels | −3.81 pp (−7.25 to −0.27) | | ds003810 | Log band-power ridge | kept | near minus random zeroing | +0.78 pp (−0.10 to +1.68) | | ds003810 | Log band-power ridge | kept | near zeroed minus all channels | −4.67 pp (−7.29 to −2.22) | | ds003810 | Log band-power ridge | kept | random zeroed minus all channels | −5.45 pp (−8.25 to −2.72) | | ds003810 | Log band-power ridge | refitted | near minus random zeroing | +0.39 pp (−1.07 to +1.80) | | ds003810 | Log band-power ridge | refitted | near zeroed minus all channels | −1.66 pp (−3.17 to −0.09) | | ds003810 | Log band-power ridge | refitted | random zeroed minus all channels | −2.04 pp (−3.95 to −0.08) | ### Limits of this block - Channels are zeroed after preprocessing, at each model’s input: not a physical montage. - Readout mismatch is a competing explanation for the fixed head’s drop, not a unique causal decomposition; the recovery does not show that the zeroed channels carried no task signal. - Near-minus-random directions differ between MAE and JEPA. - Motor imagery against rest has 10 people, and its band-power ridge starts at 57.33%, near the 50% chance level. - The log band-power ridge is a different recipe from the core Spectral ridge row. [The same comparator beside the core rows of each core protocol →](https://bci.report/protocols/arithmetic-rest/#logpower-ridge) - MAE and JEPA here are eeg-fm-masking’s published weights, not NSP weights. On all channels, MAE’s two scores and JEPA’s motor-imagery score equal the published frozen-probe cells of these checkpoints; JEPA’s EEGMAT score differs from its published cell. [Method page →](https://bci.report/methods/eeg-fm-masking/) The numeric audit replayed 90 heads and 332,640 predictions; it did not recompute encoder outputs. ### What else bounds this panel - EEGMAT’s recording order (rest before arithmetic), the source’s prior ICA cleaning and EEGMAT’s use as a development cohort are [stated in the NeuroLM panel](https://bci.report/topics/pilot-studies/#neurolm-eegmat) and bind all three blocks. - The numeric audits replayed heads, predictions and summaries from sealed features; they did not rerun the encoders or the training, and the historical raw recordings were not re-hashed. ### Not run - The paper’s own encoder: no author checkpoint was found in the sources checked (the paper, its reference code, and model and code searches). - The discriminating controls the design names: shuffled coordinates and distance-matched random context; normalisation-only, channel-only and time-only ResID; batch-partner and global-statistics controls; spectrum- and power-matched mismatch; regulariser controls; a subject-identity probe; longer or larger training and other cohorts. ### Not published here - Per-person, per-fold, per-seed and per-window values of any kind: fold and seed-by-fold results, each job’s diagnostics and their spread, participant identifiers, the selected label rows of the fitting-label subsets, features, fitted heads, predictions, checkpoints and the sealed files. Only equal-person means, paired differences with their descriptive intervals, probe and cosine-drop means over jobs, and counts are published. - Any figure or text of the NSP paper, which is cited only. - Compute, hosts, storage, paths, run times and memory. [what this release publishes, on the data-use page →](https://bci.report/data-use/#sources-nsp-components) ### Credits, licences and consent ### Neural State Prediction (NSP) [Neural State Prediction: Obstructing Shortcut Learning in EEG Foundation Models, arXiv:2609.31167v1 ↗](https://arxiv.org/abs/2609.31167v1) · [CC BY-NC-ND 4.0 ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/): cited only; no text or figure is reproduced. The local model follows the paper’s MIT-licensed reference code. ### eeg-fm-masking, the two encoders [arXiv:2609.33487 ↗](https://arxiv.org/abs/2609.33487). As run in this panel: mae-r9cm-L2 at `c861ade23dddcb0fc0be718944334a2316d635a5`, SHA-256 `d889abd87a245b92f782246f6fd51b23120208f307bc65b4de587a76ea5ed4f2`; jepa-r9cm-L2 at `88877d4cf2304cd3419181a7b1b416ce1f5d508f`, SHA-256 `e9fb1589ddc701017b7bb84035142552b9ee7d512f1dc3267fe24bc1978313b4`. Weights CC-BY-4.0 (cite arXiv:2609.33487); code MIT [↗](https://github.com/PierreGtch/eeg-fm-masking) at `38a089e63d019c23793201fff58a27d7ce399c1a`. [Method page →](https://bci.report/methods/eeg-fm-masking/) ### EEGMAT, the recordings [Credit, ethics and consent: in the NeuroLM panel’s credits above ↑](https://bci.report/topics/pilot-studies/#neurolm-credits) · [Open Data Commons Attribution License 1.0](https://opendatacommons.org/licenses/by/1-0/) · [Dataset page →](https://bci.report/datasets/eegmat/) ### ds003810, the recordings [Credit, ethics and consent: in the NeuroRVQ panel’s credits above ↑](https://bci.report/topics/pilot-studies/#neurorvq-credits) · [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset page →](https://bci.report/datasets/ds003810/) Checkpoints and weights: none is distributed; each SHA-256 identifies the file used and grants nothing. Derived analysis by BCI Report, aggregates only; not endorsed by the authors of either paper or the data’s authors, and no raw data are redistributed. Data source: [nsp-components-update.json](https://bci.report/data/nsp-components-update.json) · schema bci-report-nsp-components-update-v1. ## Related questions - Signals other than scalp EEG, human fNIRS and one macaque’s intracortical spiking: [Do simple baselines work on signals other than scalp EEG?](https://bci.report/topics/beyond-scalp-eeg/) - Frozen foundation-model encoders against matched randomly initialized ones, and further checkpoints in one frozen recipe: [Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/) - Its own page, with its frozen-encoder results beside the core matrix and its licence status: [EEGPT](https://bci.report/methods/eegpt/). - Public EEG datasets read from their own files, from the runs in a release to the windows a check can use: [What do the files show once a public EEG dataset is checked end to end?](https://bci.report/topics/data-quality/) Keep exploring Transfer - [— Sensor transfer **Dry vs. wet electrodes**](https://bci.report/topics/dry-vs-wet/) - [— Context transfer **Screen to VR**](https://bci.report/topics/screen-to-vr/) - [— Montage **Fewer electrodes**](https://bci.report/topics/fewer-electrodes/) - [— Motion robustness **On the move**](https://bci.report/topics/on-the-move/) - [— Session transfer · same person **Later sessions**](https://bci.report/topics/later-sessions/) Adapting models - [— Calibration budget **How much calibration?**](https://bci.report/topics/calibration-budget/) - [— Model adaptation **Which part to update?**](https://bci.report/topics/model-adaptation/) - [— Representation controls **Does pretraining help?**](https://bci.report/topics/does-pretraining-help/) - [— Shared encoder · route 2 **One model, several questions**](https://bci.report/topics/shared-encoder/) - [— Questions in language · route 3 **Questions in language**](https://bci.report/topics/questions-in-language/) - [— Route 3 follow-ups **New SSVEP frequencies**](https://bci.report/topics/new-ssvep-frequencies/) Reliability & clinical - [— Abstention · decision research **When not to act**](https://bci.report/topics/when-not-to-act/) - [— Sleep staging · simple baselines **Sleep-stage balance**](https://bci.report/topics/sleep-staging/) - [— Clinical research **Clinical groups**](https://bci.report/topics/clinical-groups/) - [— Data quality and reproducibility **Checked end to end**](https://bci.report/topics/data-quality/) Other signals & pilots - [— Other signals · human and macaque **Beyond scalp EEG**](https://bci.report/topics/beyond-scalp-eeg/) - [— Pilot studies · small cohorts **Pilot studies**](https://bci.report/topics/pilot-studies/) [All questions and the transfer-coverage map →](https://bci.report/topics/) ## Cite this page BCI Report (2026). *What can a small pilot show, and what can it not?* https://bci.report/topics/pilot-studies/ Figures from releases [`nsp-components-update-20261106`](https://bci.report/releases/#nsp-components-update-20261106) (2026-11-06), [`neurorvq-update-20261028`](https://bci.report/releases/#neurorvq-update-20261028) (2026-10-28), [`eegmat-ten-second-update-20261028`](https://bci.report/releases/#eegmat-ten-second-update-20261028) (2026-10-28), [`neurolm-eegmat-update-20261020`](https://bci.report/releases/#neurolm-eegmat-update-20261020) (2026-10-20), [`pilot-studies-update-20261012`](https://bci.report/releases/#pilot-studies-update-20261012) (2026-10-12). Cite the upstream datasets as well: their credits are on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/topics/pilot-studies/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Jev-style questions on EEG: the evidence Decision research · three routes Three routes, each answered by BCI Report’s own runs on public EEG data. ## What the name means Jev is [TypeSafe AI’s decision model ↗](https://typesafe.ai/blog/introducing-system-one-models-and-jev). Software sends it the state of an application and typed questions; it returns typed decisions, such as a choice from a fixed set, a score or a yes or no, each with a probability or a confidence, instead of generating text. “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. This site’s decision-research plan started from a vision paper, *[Yu & Yao, 2026 · Visual Jev: Accurate and Efficient Decisions from Shared Visual Context ↗](https://arxiv.org/abs/2609.25845)*, which applies the same encode-once, answer-many idea to images: one encoding of an image and its context serves several independent forced-choice questions. The three routes below ask whether that holds up on EEG. Three routes ## What each route found The figures, their intervals and their limits are on each route’s own page. - First route Reliable decisions **When should a decoder decline to decide?** Rescoring the saved outputs of models this site already publishes: a shared cut-off gives each method its own coverage, so methods are compared at matched coverage. There, a learned reject option did not err less than calibrated confidence for any method, and a certified risk accepted nothing on one of the two protocols. [The results, on When not to act →](https://bci.report/topics/when-not-to-act/#reliable-decisions) - Second route One representation, several questions **Can one EEG model answer several questions as well as separate models?** Fixed outputs on one shared encoder were cheaper than a head told which question it answers, but were not shown to do as well; telling the head the question, by an identifier, added nothing measurable. [One model, several questions →](https://bci.report/topics/shared-encoder/) - Third route Questions in language **Can an EEG model answer questions asked in words?** Only for questions it was trained on, and not on every dataset: in its training wording, a small head matched question numbers on seen sleep questions but lost accuracy on SSVEP; most rewordings of a seen question cost accuracy; and questions it was never trained on were not answered usefully. In a boundary probe, negated questions were answered as if they asked for what they negate. [Questions in language →](https://bci.report/topics/questions-in-language/) Fourteen follow-up experiments, thirteen of them on the same BETA cohort, tested how the question is asked, the head, phase, the matching computation, a learned correction to CCA, the cross-spectral matrix, how CCA’s statistics are pooled and weighted, and filter-bank CCA with shorter windows and fewer electrodes; five ran on a second dataset, one of them only there: [what lets a decoder answer an SSVEP frequency it never trained on →](https://bci.report/topics/new-ssvep-frequencies/) The plan, its literature and what each route set out to test: [the research plan on When not to act →](https://bci.report/topics/when-not-to-act/#decision-research) A longer write-up of the first route’s BETA results is on Hugging Face: [When should an EEG decoder abstain? ↗](https://huggingface.co/blog/Twu31/when-should-an-eeg-decoder-abstain); and of the second route’s motor-imagery results: [Can one EEG encoder answer several questions? ↗](https://huggingface.co/blog/Twu31/one-eeg-encoder-several-questions); and, in Towards AI, of the third route’s BETA results: [Can an EEG decoder answer a question it was never trained on? ↗](https://pub.towardsai.net/can-an-eeg-decoder-answer-a-question-it-was-never-trained-on-a9b0b7525b39) BCI Report is an independent project; it does not use or evaluate TypeSafe’s Jev or the Visual Jev model. --- Markdown copy of https://bci.report/jev-style/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Datasets with results on BCI Report Datasets Every published figure measured on each dataset, with its cohort, interval and source page. No raw recording is redistributed. [The open register: every EEG dataset we know of, and how far we checked it (488 rows) →](https://bci.report/datasets/register/) ## Datasets with published results | Dataset | Task | People | Appears in | Licence | | --- | --- | --- | --- | --- | | [ds003810](https://bci.report/datasets/ds003810/) | Motor imagery / rest | 10 | 7 groups | CC0-1.0 | | [EEGMAT](https://bci.report/datasets/eegmat/) | Mental arithmetic / rest | 36 | 9 groups | Open Data Commons Attribution License 1.0 | | [BETA](https://bci.report/datasets/beta/) | 40-target SSVEP | 70 | 8 groups | CC BY 4.0 | | [ds006593](https://bci.report/datasets/ds006593/) | P300 target ERP | 21 | 3 groups | CC0-1.0 | | [TMNRED / ds005383](https://bci.report/datasets/tmnred/) | Semantic target ERP | 30 | 3 groups | OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 | | [EESM19 scalp subset](https://bci.report/datasets/eesm19/) | Five-stage sleep | 20 | 4 groups | CC0-1.0 declared by upstream and mirror | | [ds005342](https://bci.report/datasets/ds005342/) | Cue-gated idle / command | 4 | 3 groups | CC0-1.0 | | [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) | 12-target SSVEP, dry and wet electrodes | 102 | 4 groups | CC BY 4.0 | | [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) | SSVEP and ERP while standing, walking and running | 17 / 23 / 24 | 3 groups | CC BY 4.0 | | [EESM23 · Ear-EEG Sleep Monitoring 2023](https://bci.report/datasets/eesm23/) | Five-stage sleep, in-ear and scalp | 10 | 1 group | CC0-1.0 | | [EEG Alpha Waves dataset](https://bci.report/datasets/alpha-waves/) | Eyes open / eyes closed | 19 | 1 group | CC-BY-4.0 | | [Phantom EEG with motion, muscle and eye artifacts](https://bci.report/datasets/phantom/) | Physical head phantom with motion, muscle and eye artifacts | — | 1 group | CC0-1.0 | | [OpenNeuro ds004584 · Parkinson's disease, rest eyes open](https://bci.report/datasets/ds004584/) | Resting-state EEG, Parkinson's disease and controls | 149 | 1 group | CC0-1.0 | | [Cattan PC/VR P300 dataset](https://bci.report/datasets/vr-pc-p300/) | P300 on a PC screen and in a VR headset | 21 | 1 group | CC-BY-4.0 | | [Multimodal gait dataset · treadmill walking](https://bci.report/datasets/gait-eeg/) | Treadmill walking at three speeds | 58 | 1 group | CC-BY-4.0 | | [YSU asynchronous SSVEP-BCI dataset](https://bci.report/datasets/ysu-async-ssvep/) | Asynchronous SSVEP, control and non-control states | 4 / 20 | 2 groups | CC-BY-4.0 | | [LTRSVP · EEG Signals from an RSVP Task](https://bci.report/datasets/ltrsvp/) | P300 target images in rapid serial visual presentation at three rates | 9 | 1 group | Open Data Commons Attribution License 1.0 | | [Dreem Open Datasets (DOD-H and DOD-O)](https://bci.report/datasets/dreem-dod/) | Five-stage sleep staging, healthy sleepers and people with obstructive sleep apnoea, kept apart | 25 / 55 | 2 groups | MIT | | [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/) | Motor imagery: left or right hand across two sessions, and imagery or rest | 51 | 2 groups | CC0-1.0 | | [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) | Overnight sleep, six polysomnography EEG channels, human-consensus stages | 100 | 2 groups | CC0-1.0 | | [WBCIC-SHU motor imagery dataset](https://bci.report/datasets/wbcic-shu/) | Motor imagery across three recording sessions: two-class and three-class cohorts, kept apart | 11 / 51 | 2 groups | CC BY 4.0 | | [Longitudinal ERP dataset (RSVP)](https://bci.report/datasets/longitudinal-rsvp/) | RSVP face-target ERP, the same people at nominal Day 1, 7, 80 and 200 visits | 15 | 1 group | CC0-1.0 | | [Forenzo & He continuous-tracking EEG-BCI dataset](https://bci.report/datasets/forenzo-continuous-tracking/) | Continuous cursor tracking with a noninvasive BCI, across sessions | — | 4 groups | CC BY 4.0 | | [Orthosis error-detection dataset (Kueper et al. 2024)](https://bci.report/datasets/orthosis-error-detection/) | Injected orthosis device errors against standard device events, EEG only: one cohort in three configurations | 8 | 3 groups | CC BY 4.0 | | [BINGO cross-linguistic inner-speech dataset](https://bci.report/datasets/bingo-imagined-speech/) | Inner speech of 15 matched nouns in Greek and in English, across people and languages | — | 1 group | No licence stated | | [EEG-fNIRS-based Handwriting Trajectory Dataset](https://bci.report/datasets/eeg-fnirs-handwriting/) | Four released coded labels in an imagined-handwriting competition dataset, EEG only: a data-quality and methods case study | 5 | 1 group | CC BY-NC-SA 4.0 | | [REFED emotion dataset (EEG-fNIRS)](https://bci.report/datasets/refed/) | Self-reported valence and arousal while watching videos: a stored-clip diagnostic | — | 1 group | CC BY-NC-SA 4.0 | | [Bak et al. fNIRS finger- and foot-tapping dataset](https://bci.report/datasets/bak-fnirs/) | Human fNIRS: right-hand, left-hand and foot tapping, a person left out at a time | 30 | 1 group | CC BY 4.0 | | [MC_RTT (Neural Latents Benchmark)](https://bci.report/datasets/nlb-mc-rtt/) | One macaque’s motor-cortex spiking during self-paced reaching: finger velocity | — | 1 group | CC BY 4.0 | | [OpenNeuro ds007420 · multi-distance fNIRS](https://bci.report/datasets/ds007420/) | Human fNIRS recordings: a numerical coverage check, no decoding | 12 | 1 group | CC0-1.0 | | [Shin2022 · closed-loop motor imagery EEG (Figshare 20383716)](https://bci.report/datasets/shin2022/) | Closed-loop cursor control: the instructed right or left target, dry electrodes | 10 | 1 group | CC BY 4.0 | | [Kim beta-range SSVEP · 40 targets, 14.0–21.8 Hz (Figshare 28806815)](https://bci.report/datasets/kim2025-beta-range/) | 40-target SSVEP speller, 14.0 to 21.8 Hz | 40 | 1 group | CC BY 4.0 | BOAS is published with three stated gaps: its consent statement does not say whether participants agreed to public sharing or secondary use; its ethics and consent statements come only from the publisher, and no peer-reviewed paper describes it; its ethics reference was added in version 1.1.1 with no approval date. [Its page](https://bci.report/datasets/boas/) states them in full, with the credit. ## Datasets with results in a section of a question page These have no page of their own; each row links the section that publishes its results. | Dataset | Task | Results | Licence | | --- | --- | --- | --- | | [IESSEEG ↗](https://huggingface.co/datasets/Capur/IESSEEG/tree/1c32ee3a425a324808a9d275b60ce85625af35c8) | Case/control discrimination from one awake minute of EEG per person | [Clinical groups: IESSEEG](https://bci.report/topics/clinical-groups/#iesseeg) | CC BY 4.0 | | [MIND ↗](https://doi.org/10.57760/sciencedb.34326) | Four-class directional motor imagery of the right upper limb with EEG and fNIRS; here the EEG of stored codes 4 versus 5 and 6 versus 7. | [Data quality: MIND](https://bci.report/topics/data-quality/#mind2026) | CC BY 4.0 · CC BY-NC 4.0, as its sources state them | | [eldBETA ↗](https://doi.org/10.6084/m9.figshare.18032669.v3) | Nine-target SSVEP selection by older adults, recorded in seven blocks per person. | [Data quality: eldBETA](https://bci.report/topics/data-quality/#eldbeta) | CC BY 4.0 | | [ASEEG ↗](https://doi.org/10.5281/zenodo.18029536) | Case/control discrimination from resting EEG, beside age and sex | [Clinical groups: ASEEG](https://bci.report/topics/clinical-groups/#aseeg) | CC BY 4.0 | ## Listed, not scored Sources reviewed or considered for the core matrix that carry no result in any release. - [BNCI2014-008 ↗](https://bnci-horizon-2020.eu/database/data-sets/008-2014) · Research source · Not in this release - [BNCI2014-009 ↗](https://bnci-horizon-2020.eu/database/data-sets/009-2014) · Research source · Not in this release - [stew-monster-processed ↗](https://ieee-dataport.org/open-access/simultaneous-task-eeg-workload-dataset) · Research source · Not in this release - [physionet-eegmmidb-1.0.0 ↗](https://physionet.org/content/eegmmidb/1.0.0/) · Research source · Not in this release - [BNCI2014-001 ↗](https://bnci-horizon-2020.eu/database/data-sets/001-2014) · Research source · Not in this release - [BNCI2014-004 ↗](https://bnci-horizon-2020.eu/database/data-sets/004-2014) · Research source · Not in this release [All methods →](https://bci.report/methods/) Register ## Public-data register: documented experiments Source terms are checked separately from numerical results. Motor imagery / rest ### [ds003810](https://bci.report/datasets/ds003810/) Clear CC0 snapshot plus dataset-specific ethics and signed-consent evidence; aggregate-only publication materially limits privacy exposure. Rights reviewed 2026-09-20 - Subjects: 10 - Channels: 15 Aggregate results — [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ Arithmetic / rest ### [EEGMAT](https://bci.report/datasets/eegmat/) Attribution license and study-specific approval/consent are documented; eligibility assumes aggregate-only output and exclusion of subject-info fields. Rights reviewed 2026-09-20 - Subjects: 36 - Channels: 19 Aggregate results — [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. Five-stage sleep ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) Clear upstream CC0 and study ethics evidence; use only the pinned scalp-signal derivative and publish aggregates, with provenance disclosed. Consent covered the study, not publication; the public release rests on a GDPR anonymization assessment (amended 2026-09-22). Rights reviewed 2026-09-20 - Subjects: 20 - Channels: 6 Aggregate results — [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) — Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. 40-target SSVEP ### [BETA](https://bci.report/datasets/beta/) Eligible only under the stated personal, research-led, noncommercial operation. Any ads, sponsorship, fees, or use directed toward commercial advantage requires fresh review or permission. Rights reviewed 2026-09-20 - Subjects: 70 - Channels: 4 / 8 selected Aggregate results — [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. Research source ### [BNCI2014-008 ↗](https://bnci-horizon-2020.eu/database/data-sets/008-2014) Small clinical cohort and linked health attributes create disproportionate reidentification risk; aggregate-only output does not resolve the missing secondary-publication consent evidence. Rights reviewed 2026-09-20 - Subjects: — - Channels: — Not in this release — [CC BY-NC-ND 4.0 ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/) — BNCI Horizon 2020 catalog · Original researchers credited at source. Research source ### [BNCI2014-009 ↗](https://bnci-horizon-2020.eu/database/data-sets/009-2014) License permits only noncommercial unadapted redistribution, and the consent/ethics chain for public secondary results remains unverified. Rights reviewed 2026-09-20 - Subjects: — - Channels: — Not in this release — [CC BY-NC-ND 4.0 ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/) — BNCI Horizon 2020 catalog · Original researchers credited at source. Research source ### [stew-monster-processed ↗](https://ieee-dataport.org/open-access/simultaneous-task-eeg-workload-dataset) Ethics evidence is positive, but upstream license and the processed mirror's authority to apply CC BY remain unresolved. Rights reviewed 2026-09-20 - Subjects: — - Channels: — Not in this release — [Mirror declares CC BY 4.0; exact upstream DataPort license not independently verified ↗](https://creativecommons.org/licenses/by/4.0/) — monster-monash/STEW Hugging Face processed mirror · Original researchers credited at source. Semantic target ERP ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) Strong study-specific open-sharing evidence. Aggregate metrics avoid redistribution of stimulus text and participant metadata; credit under CC BY. Rights reviewed 2026-09-20 - Subjects: 30 - Channels: 30 Aggregate results — [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0. P300 target ERP ### [ds006593](https://bci.report/datasets/ds006593/) Pinned CC0 release with dataset-specific IRB and consent evidence and comparatively low metadata exposure. Rights reviewed 2026-09-20 - Subjects: 21 - Channels: 19 Aggregate results — [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) — OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. Research source ### [physionet-eegmmidb-1.0.0 ↗](https://physionet.org/content/eegmmidb/1.0.0/) Reuse license is clear, but the dataset-specific human-subject consent and ethics chain is not evidenced strongly enough for a new public benchmark release. Rights reviewed 2026-09-20 - Subjects: — - Channels: — Not in this release — [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) — PhysioNet EEG Motor Movement/Imagery Dataset 1.0.0 · Original researchers credited at source. Research source ### [BNCI2014-001 ↗](https://bnci-horizon-2020.eu/database/data-sets/001-2014) Per-dataset license is verified, but consent evidence and the application of ND to the planned publication pipeline should be clarified. Rights reviewed 2026-09-20 - Subjects: — - Channels: — Not in this release — [CC BY-ND 4.0 ↗](https://creativecommons.org/licenses/by-nd/4.0/) — BNCI Horizon 2020 catalog · Original researchers credited at source. Research source ### [BNCI2014-004 ↗](https://bnci-horizon-2020.eu/database/data-sets/004-2014) Per-dataset license is verified, but consent evidence and the application of ND to the planned publication pipeline should be clarified. Rights reviewed 2026-09-20 - Subjects: — - Channels: — Not in this release — [CC BY-ND 4.0 ↗](https://creativecommons.org/licenses/by-nd/4.0/) — BNCI Horizon 2020 catalog · Original researchers credited at source. Cue-gated idle / command ### [ds005342](https://bci.report/datasets/ds005342/) Eligible only as an aggregate-only case study: remove all participant IDs/results and condition breakdowns, label n=4 prominently, avoid subgroup claims, and make no claim that the cohort is anonymous. Rights reviewed 2026-09-20 - Subjects: 4 evaluated / 32 prepared - Channels: 17 Aggregate results — [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) — OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. --- Markdown copy of https://bci.report/datasets/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEG Alpha Waves dataset: EEG decoding results Eyes open / eyes closed Resting-state eyes-open/eyes-closed EEG from 20 participants, recorded with a 16-electrode headset during a 2017 pilot experiment at GIPSA-lab, Grenoble, by Cattan, Rodrigues and Congedo. Released on Zenodo in 2018 (CC BY 4.0) with a research report (hal-02086581); dataset ID ALPHA.EEG.2017-GIPSA. **Also known as** EEG Alpha Waves dataset · ALPHA.EEG.2017-GIPSA · Rodrigues2017 · Zenodo 2605110 · hal-02086581 Description sources [zenodo.org](https://zenodo.org/records/2605110) · [hal.science](https://hal.science/hal-02086581) · [moabb.neurotechx.com](https://moabb.neurotechx.com/docs/generated/moabb.datasets.Rodrigues2017.html) ## Where it appears - [Can fewer electrodes, or electrodes in the ear, match a full scalp montage?](https://bci.report/topics/fewer-electrodes/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Fewer electrodes · four posterior vs. all sixteen Source page: [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) · Chance level 50.0% · [evidence-update.json](https://bci.report/data/evidence-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Relative band-power logistic regression · Four posterior electrodes: 79.5% (72.6%–85.8%) - Relative band-power logistic regression · All sixteen electrodes: 77.9% (70.5%–85.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Relative band-power logistic regression | Four posterior electrodes | Balanced accuracy | 79.5% (72.6%–85.8%) | 19 | | Relative band-power logistic regression | All sixteen electrodes | Balanced accuracy | 77.9% (70.5%–85.3%) | 19 | | Relative band-power logistic regression | All sixteen minus four posterior | Paired difference | −1.6 pp (−10.0 pp–+6.8 pp) | 19 | ## Source and licence ### EEG Alpha Waves dataset **Credit** Grégoire Cattan, Pedro L. C. Rodrigues and Marco Congedo · EEG Alpha Waves dataset, Zenodo, doi:10.5281/zenodo.2605110. Report: Grégoire Cattan, Pedro Luiz Coelho Rodrigues and Marco Congedo · EEG Alpha Waves Dataset, research report, GIPSA-lab, Grenoble, 2018, hal-02086581. [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.5281/zenodo.2605110) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-alpha-waves) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *EEG Alpha Waves dataset: EEG decoding results*. https://bci.report/datasets/alpha-waves/ Figures from release [`evidence-update-20260922`](https://bci.report/releases/#evidence-update-20260922) (2026-09-22). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/alpha-waves/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Bak et al. fNIRS finger- and foot-tapping dataset: fNIRS results Human fNIRS: right-hand, left-hand and foot tapping, a person left out at a time The open-access fNIRS dataset for classification of unilateral finger- and foot-tapping (figshare 9783755) holds functional near-infrared spectroscopy from 30 volunteers performing right-hand finger tapping, left-hand finger tapping and foot tapping, 25 trials of each, recorded with a three-wavelength multichannel system (LIGHTNIRS, Shimadzu) with eight sources and eight detectors placed around C3 and C4. Released by Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong (Korea University and Wonkwang University) under CC BY 4.0 and described in Electronics (2019). The supplied oxy- and deoxyhemoglobin series of all 30 participants are used here. **Also known as** Open Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping · figshare 9783755 · fNIRS dataset for ternary classification · Bak et al. (2019) Description sources [doi.org](https://doi.org/10.6084/m9.figshare.9783755.v2) · [doi.org](https://doi.org/10.3390/electronics8121486) · [api.figshare.com](https://api.figshare.com/v2/articles/9783755/versions/2) ## Where it appears - [Do simple baselines work on signals other than scalp EEG?](https://bci.report/topics/beyond-scalp-eeg/#bak-fnirs) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Beyond scalp EEG · human fNIRS, right-hand, left-hand and foot tapping Source page: [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/#bak-fnirs) · Chance level 33.3% · [beyond-scalp-update.json](https://bci.report/data/beyond-scalp-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Standardization + shrinkage LDA | A person left out of training · executed tapping, not motor imagery | Balanced accuracy | 66.8% (60.6%–72.9%) — Each person’s balanced accuracy, averaged with equal weight. Class recall over all held-out trials: right hand 69.2%, left hand 63.6%, foot 67.7%. The interval resamples the fixed participant scores (2,000 draws): not a refit bootstrap and not a clinical confidence interval. | 30 | | Training prior — comparator | A person left out of training · executed tapping, not motor imagery | Balanced accuracy | 33.3% — Predicts the same class every time: no interval, because every person gets the same score. | 30 | Offline, from the supplied oxy- and deoxyhemoglobin series over a window after each task onset. Task-correlated systemic or motion effects are not excluded, so this shows neither a brain-only signal nor cortical localisation. This site’s own protocol: it does not reproduce the authors’ within-person scores, and it is not compared with any scalp EEG result. ## Source and licence ### Bak et al. fNIRS finger- and foot-tapping dataset **Credit** Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong · Open Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, figshare, doi:10.6084/m9.figshare.9783755.v2, CC BY 4.0. Paper: Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong, Open-Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, Electronics 8(12), 1486 (2019), doi:10.3390/electronics8121486. Derived analysis by BCI Report; not endorsed by the authors. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.6084/m9.figshare.9783755.v2) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-bak-fnirs) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Bak et al. fNIRS finger- and foot-tapping dataset: fNIRS results*. https://bci.report/datasets/bak-fnirs/ Figures from release [`beyond-scalp-update-20261012`](https://bci.report/releases/#beyond-scalp-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/bak-fnirs/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # BETA: EEG decoding results 40-target SSVEP BETA (BEnchmark database Towards BCI Application) is a 40-target SSVEP dataset: 64-channel EEG from 70 participants performing a cued-spelling task on a keyboard-style virtual speller (8–15.8 Hz), four blocks each, recorded outside an electromagnetically shielded room. Released by Liu et al. (Tsinghua University) in Frontiers in Neuroscience (2020), with data on figshare. **Also known as** BETA · BETA database · The BETA database · BETA: A Large Benchmark Database Toward SSVEP-BCI Application · Liu et al. (2020) Description sources [figshare.com](https://figshare.com/articles/dataset/_/12264401) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC7324867/) · [doi.org](https://doi.org/10.3389/fnins.2020.00627) · [bci.med.tsinghua.edu.cn](https://bci.med.tsinghua.edu.cn/download.html) ## Where it appears - [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) - [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) - [How often does an EEG decoder fire when nobody is giving a command?](https://bci.report/topics/when-not-to-act/) - [Can an EEG model answer questions asked in words?](https://bci.report/topics/questions-in-language/) Follow-up tests on BETA’s held-out frequencies: [What lets a decoder answer an SSVEP frequency it never trained on?](https://bci.report/topics/new-ssvep-frequencies/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-beta-8ch.csv](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) **Balanced accuracy, 8 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Standard CCA: 63.1% (57.2%–69.0%) - Spectral ridge: 50.7% (45.6%–55.8%) - LaBraM: 10.8% (9.5%–12.2%) - CBraMod: 33.7% (29.8%–37.6%) - EEGNet: 55.8% (50.1%–61.3%) - Filter-bank CCA (FBCCA-M3-N7-H5): 79.1% (75.1%–82.8%) - ShallowFBCSPNet: 23.0% (20.3%–25.9%) - Deep4Net: 52.8% (47.3%–58.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 63.1% (57.2%–69.0%) | 70 | | Spectral ridge | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 50.7% (45.6%–55.8%) | 70 | | [LaBraM](https://bci.report/methods/labram/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 10.8% (9.5%–12.2%) | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 33.7% (29.8%–37.6%) | 70 | | [EEGNet](https://bci.report/methods/eegnet/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 55.8% (50.1%–61.3%) | 70 | | [Filter-bank CCA (FBCCA-M3-N7-H5)](https://bci.report/methods/fbcca/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 79.1% (75.1%–82.8%) — FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free · 250 Hz | 70 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 23.0% (20.3%–25.9%) — One fixed seed and training budget, as the EEGNet cell | 70 | | [Deep4Net](https://bci.report/methods/deep4net/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 52.8% (47.3%–58.2%) — One fixed seed and training budget, as the EEGNet cell | 70 | ### Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-beta-4ch.csv](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) **Balanced accuracy, 8 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Standard CCA: 57.6% (51.8%–63.5%) - Spectral ridge: 48.2% (43.1%–53.3%) - LaBraM: 12.9% (11.2%–14.6%) - CBraMod: 27.6% (24.1%–31.3%) - EEGNet: 44.1% (38.6%–49.6%) - Filter-bank CCA (FBCCA-M3-N7-H5): 72.3% (67.5%–76.9%) - ShallowFBCSPNet: 16.2% (14.1%–18.6%) - Deep4Net: 38.7% (33.7%–43.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 57.6% (51.8%–63.5%) | 70 | | Spectral ridge | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 48.2% (43.1%–53.3%) | 70 | | [LaBraM](https://bci.report/methods/labram/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 12.9% (11.2%–14.6%) | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 27.6% (24.1%–31.3%) | 70 | | [EEGNet](https://bci.report/methods/eegnet/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 44.1% (38.6%–49.6%) | 70 | | [Filter-bank CCA (FBCCA-M3-N7-H5)](https://bci.report/methods/fbcca/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 72.3% (67.5%–76.9%) — FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free · 250 Hz | 70 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 16.2% (14.1%–18.6%) — One fixed seed and training budget, as the EEGNet cell | 70 | | [Deep4Net](https://bci.report/methods/deep4net/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 38.7% (33.7%–43.8%) — One fixed seed and training budget, as the EEGNet cell | 70 | ### Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) **Balanced accuracy, 16 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 37.5% (33.3%–41.9%) - REVE Large · Frozen encoder + ridge head: 52.7% (47.7%–57.8%) - LUNA Base · Frozen encoder + ridge head: 8.6% (7.7%–9.5%) - LUNA Large · Frozen encoder + ridge head: 10.0% (8.9%–11.1%) - BrainOmni Base · Frozen encoder + ridge head: 21.7% (18.6%–25.0%) - CodeBrain · Frozen encoder + ridge head: 11.9% (10.5%–13.3%) - EEGMamba · Frozen encoder + ridge head: 40.2% (35.7%–44.7%) - ST-EEGFormer Base · Frozen encoder + ridge head: 17.1% (15.5%–18.8%) - ST-EEGFormer Large · Frozen encoder + ridge head: 22.6% (20.1%–25.3%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 54.8% (49.4%–60.0%) - ERP-FM Base · Frozen encoder + ridge head: 18.4% (16.3%–20.5%) - SingLEM · Frozen encoder + ridge head: 26.6% (23.7%–29.5%) - ZUNA 1.1 · Frozen encoder + ridge head: 11.9% (10.8%–13.0%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 49.6% (44.5%–54.8%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 53.4% (48.1%–58.6%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 51.0% (45.8%–56.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 37.5% (33.3%–41.9%) | 70 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 52.7% (47.7%–57.8%) | 70 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 8.6% (7.7%–9.5%) | 70 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 10.0% (8.9%–11.1%) | 70 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 21.7% (18.6%–25.0%) | 70 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 11.9% (10.5%–13.3%) | 70 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 40.2% (35.7%–44.7%) | 70 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 17.1% (15.5%–18.8%) — Pretraining exposure: in the authors' published pretraining list. | 70 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 22.6% (20.1%–25.3%) — Pretraining exposure: in the authors' published pretraining list. | 70 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 54.8% (49.4%–60.0%) | 70 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 18.4% (16.3%–20.5%) | 70 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 26.6% (23.7%–29.5%) | 70 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 11.9% (10.8%–13.0%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 70 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 49.6% (44.5%–54.8%) | 70 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 53.4% (48.1%–58.6%) | 70 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 51.0% (45.8%–56.1%) | 70 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) **Balanced accuracy, 16 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 25.5% (22.4%–28.8%) - REVE Large · Frozen encoder + ridge head: 46.4% (41.7%–51.2%) - LUNA Base · Frozen encoder + ridge head: 9.0% (8.1%–10.1%) - LUNA Large · Frozen encoder + ridge head: 11.2% (9.8%–12.6%) - BrainOmni Base · Frozen encoder + ridge head: 13.0% (10.6%–15.6%) - CodeBrain · Frozen encoder + ridge head: 11.2% (9.8%–12.7%) - EEGMamba · Frozen encoder + ridge head: 32.7% (28.9%–36.7%) - ST-EEGFormer Base · Frozen encoder + ridge head: 17.5% (15.8%–19.3%) - ST-EEGFormer Large · Frozen encoder + ridge head: 22.4% (19.9%–25.0%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 47.1% (42.1%–52.1%) - ERP-FM Base · Frozen encoder + ridge head: 20.9% (18.4%–23.5%) - SingLEM · Frozen encoder + ridge head: 20.5% (17.8%–23.3%) - ZUNA 1.1 · Frozen encoder + ridge head: 8.4% (7.5%–9.4%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 43.4% (38.7%–48.3%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 46.7% (41.9%–51.5%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 45.4% (40.6%–50.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 25.5% (22.4%–28.8%) | 70 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 46.4% (41.7%–51.2%) | 70 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 9.0% (8.1%–10.1%) | 70 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 11.2% (9.8%–12.6%) | 70 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 13.0% (10.6%–15.6%) | 70 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 11.2% (9.8%–12.7%) | 70 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 32.7% (28.9%–36.7%) | 70 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 17.5% (15.8%–19.3%) — Pretraining exposure: in the authors' published pretraining list. | 70 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 22.4% (19.9%–25.0%) — Pretraining exposure: in the authors' published pretraining list. | 70 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 47.1% (42.1%–52.1%) | 70 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 20.9% (18.4%–23.5%) | 70 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 20.5% (17.8%–23.3%) | 70 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 8.4% (7.5%–9.4%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 70 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 43.4% (38.7%–48.3%) | 70 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 46.7% (41.9%–51.5%) | 70 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 45.4% (40.6%–50.2%) | 70 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### EEGPT, frozen encoder · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#eegpt) · Chance level 2.5% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 20.3% (18.2%–22.6%) | 70 | PO5 and PO6 use channel embeddings that EEGPT’s pretraining never trained: they are two of the four names in its 62-name channel dictionary (AF7, AF8, PO5, PO6) that its 58-electrode pretraining montage leaves out. The other six channels here, and all four of the 4-channel subset, use trained embeddings. For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/beta-8ch/#eegpt-terms) ### EEGPT, frozen encoder · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#eegpt) · Chance level 2.5% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 17.3% (15.5%–19.2%) | 70 | ### Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) | Fixed threshold: calibrated confidence at least 0.8 | Trials accepted | 3,176 / 11,200 — 0 of 70 people had nothing accepted. | 70 | | [Standard CCA](https://bci.report/methods/cca/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.3 pp (−0.3 pp–+1.0 pp) — No difference resolved: the interval includes zero. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | Fixed threshold: calibrated confidence at least 0.8 | Trials accepted | 165 / 11,200 — 38 of 70 people had nothing accepted. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | −0.03 pp (−0.8 pp–+0.6 pp) — No difference resolved: the interval includes zero. | 70 | | [EEGNet](https://bci.report/methods/eegnet/) | Fixed threshold: calibrated confidence at least 0.8 | Trials accepted | 1,863 / 11,200 — 10 of 70 people had nothing accepted. | 70 | | [EEGNet](https://bci.report/methods/eegnet/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.8 pp (+0.02 pp–+1.6 pp) — Difference resolved: the interval excludes zero. | 70 | **Limits:** Uniform target prior by design; cue-paced lab spelling with no idle state, so no false-activation figure of any kind. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### Questions in language · SSVEP (route 3) Source page: [Questions in language](https://bci.report/topics/questions-in-language/) · [questions-in-language-update.json](https://bci.report/data/questions-in-language-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Plain spectrum (L0) | P1 · label template − question number (TPL − ID) · seen, 32-way · three seeds | Paired difference, language arm minus reference (pp) | −5.96 pp (−6.84–−5.08 pp) — ID higher; 2 pp margin not met; passes the triviality canary. | 70 | | Plain spectrum (L0) | P1 · description − question number (DESC − ID) · seen, 32-way · three seeds | Paired difference, language arm minus reference (pp) | −5.46 pp (−6.26–−4.66 pp) — ID higher; 2 pp margin not met; passes the triviality canary. | 70 | | Plain spectrum (L0) | P2 · new sentence frame − training wording (TPL(a1) − TPL(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −4.01 pp (−6.53–−2.12 pp) — TPL(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | Plain spectrum (L0) | P2 · synonym − training wording (TPL(a2) − TPL(train)) · seen, reworded, for these three synonyms · three seeds | Paired difference, language arm minus reference (pp) | −1.84 pp (−2.10–−1.57 pp) — TPL(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | Plain spectrum (L0) | P2 · new description − training descriptions (DESC(a3) − DESC(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −4.13 pp (−5.58–−3.11 pp) — DESC(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | Plain spectrum (L0) | P4 · label template − CCA, no training (TPL − CCA) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −52.21 pp (−55.13–−48.99 pp) — Unseen question. CCA higher; 2 pp margin not met; headroom gate passed. | 70 | | Plain spectrum (L0) | P4 · label template − average of the two neighbouring seen frequencies (TPL − NN(TPL)) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −10.98 pp (−12.24–−9.67 pp) — Unseen question. NN(TPL) higher; difference only: no margin; headroom gate passed. | 70 | | Plain spectrum (L0) | P4 · label template − numeric code (TPL − NUM) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −12.77 pp (−14.52–−10.99 pp) — Unseen question. NUM higher; difference only: no margin; headroom gate passed. | 70 | | Plain spectrum (L0) | P4 · label template − CCA, no training (TPL − CCA) · unseen frequency, neighbour 3-way · three seeds | Paired difference, language arm minus reference (pp) | −62.36 pp (−65.87–−58.69 pp) — Unseen question. CCA higher; 2 pp margin not met; headroom gate passed. | 70 | | Plain spectrum (L0) | P4 · label template − numeric code (TPL − NUM) · unseen frequency, neighbour 3-way · three seeds | Paired difference, language arm minus reference (pp) | −1.24 pp (−3.17–+0.85 pp) — Unseen question. No difference shown; difference only: no margin; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · label template − question number (TPL − ID) · seen, 32-way · three seeds | Paired difference, language arm minus reference (pp) | −3.49 pp (−4.10–−2.93 pp) — ID higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · description − question number (DESC − ID) · seen, 32-way · three seeds | Paired difference, language arm minus reference (pp) | −3.37 pp (−3.98–−2.80 pp) — ID higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new sentence frame − training wording (TPL(a1) − TPL(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −2.72 pp (−4.43–−1.43 pp) — TPL(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · synonym − training wording (TPL(a2) − TPL(train)) · seen, reworded, for these three synonyms · three seeds | Paired difference, language arm minus reference (pp) | −1.33 pp (−1.53–−1.14 pp) — TPL(train) higher; equivalent within ±2 pp; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new description − training descriptions (DESC(a3) − DESC(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −2.95 pp (−3.81–−2.26 pp) — DESC(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − CCA, no training (TPL − CCA) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −61.94 pp (−65.65–−58.06 pp) — Unseen question. CCA higher; 2 pp margin not met; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − average of the two neighbouring seen frequencies (TPL − NN(TPL)) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −1.16 pp (−1.97–−0.37 pp) — Unseen question. NN(TPL) higher; difference only: no margin; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − numeric code (TPL − NUM) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −2.46 pp (−3.23–−1.69 pp) — Unseen question. NUM higher; difference only: no margin; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − CCA, no training (TPL − CCA) · unseen frequency, neighbour 3-way (frequency and stimulus phase) · three seeds | Paired difference, language arm minus reference (pp) | −54.67 pp (−57.77–−51.47 pp) — Unseen question. CCA higher; 2 pp margin not met; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − numeric code (TPL − NUM) · unseen frequency, neighbour 3-way (frequency and stimulus phase) · three seeds | Paired difference, language arm minus reference (pp) | +15.61 pp (+14.25–+16.93 pp) — Unseen question. TPL higher; difference only: no margin; headroom gate passed. | 70 | **Limits:** 'Unseen' means unseen by the EEG head; the text encoder has read these words, and every token of a held-out frequency also occurs in the training texts. Held-out SSVEP questions name a frequency, which CCA answers without training; interpolation tested, extrapolation secondary. In the neighbour 3-way the unseen frequency competes against trained detectors for its neighbours; at L1 that test also involves stimulus phase. Rewording and unseen results are separate rows and never pooled. 35 pre-declared comparisons, not corrected for multiplicity. The pre-run permutation canary failed twice (revisions 3 and 4) and was revised by the owner, who overrode revision 4's "no further revision"; revision 5 passed. Both failed records are kept and disclosed. [Every limitation of route 3 →](https://bci.report/topics/questions-in-language/#methods-and-limits) ## Source and licence ### BETA **Credit** Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-beta) A longer write-up of the core-matrix results on BETA is on Hugging Face: [CCA and frozen EEG foundation models on BETA ↗](https://huggingface.co/blog/Twu31/cca-and-frozen-eeg-foundation-models-on-beta) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *BETA: EEG decoding results*. https://bci.report/datasets/beta/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`questions-in-language-update-20261008`](https://bci.report/releases/#questions-in-language-update-20261008) (2026-10-08), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/beta/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # BINGO cross-linguistic inner-speech dataset: EEG decoding results Inner speech of 15 matched nouns in Greek and in English, across people and languages The cross-linguistic inner-speech dataset of the BINGO (Brain Imagined-Speech Communication) project at CERTH is one of the two EEG datasets the project releases on its website. On each trial a participant saw one of 15 matched nouns, in Greek or in English, held it in memory through a countdown, and then said it silently during a 1.5-second inner-speech window; each item was repeated ten times, in fully randomised order. The project's public deliverable D3.1 describes the protocol and states that the same preprocessing, including artifact removal, was applied to both released datasets. The source states no licence. Only its complete subject records are used here. **Also known as** BINGO (Brain Imagined-Speech Communication) · BINGO cross-linguistic inner speech elicitation dataset · BINGO Dataset 2 · D3.1 – The BINGO Benchmarking Framework Description sources [bingo-project.gr](https://bingo-project.gr/datasets/) · [bingo-project.gr](https://bingo-project.gr/download/397) ## Where it appears - [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/#bingo) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Pilot studies · inner speech in Greek and English, across people and across languages (a negative result) Source page: [Pilot studies](https://bci.report/topics/pilot-studies/#bingo) · Chance level 6.7% · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Relative band power + ridge classifier · Trained on Greek, tested on Greek · a subject field left out: 6.4% (5.8%–6.9%) - Relative band power + ridge classifier · Trained on Greek, tested on English · a subject field left out: 6.6% (5.5%–7.6%) - Relative band power + ridge classifier · Trained on English, tested on English · a subject field left out: 6.7% (5.8%–7.6%) - Relative band power + ridge classifier · Trained on English, tested on Greek · a subject field left out: 7.1% (6.2%–7.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Relative band power + ridge classifier | Trained on Greek, tested on Greek · a subject field left out | Balanced accuracy | 6.4% (5.8%–6.9%) — Top-3 accuracy 19.1%, against its prior of 20.0%. | 17 subject fields | | Relative band power + ridge classifier | Trained on Greek, tested on English · a subject field left out | Balanced accuracy | 6.6% (5.5%–7.6%) — Top-3 accuracy 21.0%, against its prior of 20.0%. | 17 subject fields | | Relative band power + ridge classifier | Trained on English, tested on English · a subject field left out | Balanced accuracy | 6.7% (5.8%–7.6%) — Top-3 accuracy 19.9%, against its prior of 20.0%. | 17 subject fields | | Relative band power + ridge classifier | Trained on English, tested on Greek · a subject field left out | Balanced accuracy | 7.1% (6.2%–7.9%) — Top-3 accuracy 19.6%, against its prior of 20.0%. | 17 subject fields | | Relative band power + ridge classifier | Across languages minus within, the mean over both training languages | Paired difference | +0.3 pp (−0.6–+1.3 pp) — The interval spans zero: no superiority or equivalence between languages, and no significance test. | 17 subject fields | Subject fields are counted as the source defines them; they are not proven to be distinct people. Every direction is close to the 15-class prior: a negative result for this one baseline in this setting, with other methods untested. The intervals are approximate and descriptive, from overlapping leave-one-field-out fits. Delivered signals, as the project describes them (BINGO deliverable D3.1, version 2.0): the same preprocessing for both of its released datasets — band-pass and notch filtering, Artifact Subspace Reconstruction (ASR) and FORCe removal of ocular and muscle artifacts — and then segmentation into trials. The signals analysed here were therefore artifact-cleaned on the continuous recordings before any trial was cut. As the project states it; not re-verified against the delivered file. ## Source and licence ### BINGO cross-linguistic inner-speech dataset **Credit** BINGO (Brain Imagined-Speech Communication) project, CERTH · cross-linguistic inner-speech EEG dataset, https://bingo-project.gr/datasets/; no licence is stated by the source. Documentation: Fotis P. Kalaganis, Kostas Georgiadis, Maria Kyrou, Spiros Nikolopoulos and Ioannis Kompatsiaris, D3.1 – The BINGO Benchmarking Framework, version 2.0, public deliverable (2025), https://bingo-project.gr/download/397. Derived analysis by BCI Report; not endorsed by the authors. No licence stated · [Dataset record ↗](https://bingo-project.gr/datasets/) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report from data whose source states no licence; no right to reuse the data is implied. ### Ethics and consent, as the source states them Ethics approval: No ethics-approval reference is published by the source. Informed consent: The project’s public deliverable D3.1, version 3.0, states that each participant signed a consent form before the sessions. It says so for the project’s 20-participant NATO-alphabet dataset, recorded in the same CERTH laboratory; the cross-language dataset’s own documentation, D3.1 version 2.0, has no ethics or consent statement. The owner decided to publish these aggregates on that basis. Consent statement read from [D3.1 – The BINGO Benchmarking Framework, version 3.0 ↗](https://bingo-project.gr/download/402). [BINGO’s rights on the data-use page →](https://bci.report/data-use/#bingo-rights) [Row in the open register →](https://bci.report/datasets/register/#r-site-bingo-imagined-speech) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *BINGO cross-linguistic inner-speech dataset: EEG decoding results*. https://bci.report/datasets/bingo-imagined-speech/ Figures from release [`pilot-studies-update-20261012`](https://bci.report/releases/#pilot-studies-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/bingo-imagined-speech/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # BOAS · Bitbrain Open Access Sleep dataset: EEG decoding results Overnight sleep, six polysomnography EEG channels, human-consensus stages BOAS, the Bitbrain Open Access Sleep dataset (OpenNeuro ds005555), holds 128 full-night recordings from 100 adults of the general population, each recorded with clinical polysomnography and a Bitbrain forehead headband at the same time. Sleep stages were scored by three human experts, a fourth breaking three-way ties, and separately by the publisher's automatic model. Released by Bitbrain Technologies (Zaragoza, Spain) and colleagues under CC0; version 1.1.3 is used here. No peer-reviewed paper describes the dataset. Only the six polysomnography EEG channels and the human-consensus stages are used here. **Also known as** BOAS · Bitbrain Open Access Sleep dataset · The Bitbrain Open Access Sleep (BOAS) dataset · OpenNeuro ds005555 · ds005555 Description sources [openneuro.org](https://openneuro.org/datasets/ds005555/versions/1.1.3) · [api.datacite.org](https://api.datacite.org/dois/10.18112/openneuro.ds005555.v1.1.3) · [www.bitbrain.com](https://www.bitbrain.com/science/eeg-datasets) ## Where it appears - [Can one EEG model answer several questions about the same data as well as separate models, and at what cost?](https://bci.report/topics/shared-encoder/) - [Can an EEG model answer questions asked in words?](https://bci.report/topics/questions-in-language/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### One model, several questions · sleep (route 2) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 34.4% (32.1%–37.0%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 42.9% (40.1%–45.7%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 41.4% (38.6%–44.2%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 35.3% (32.9%–37.8%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +6.96 pp (+4.92–+9.09 pp) — C1 higher; 2 pp margin not met; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | −1.47 pp (−3.79–+0.94 pp) — No difference shown; B-sh non-inferior at 2 pp; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-A)) | Paired difference, balanced accuracy | −0.84 pp (−2.73–+0.91 pp) — No difference shown; 2 pp margin not met: inconclusive at this sample size; gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds | Balanced accuracy | 70.7% (68.5%–72.7%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds | Balanced accuracy | 71.4% (69.1%–73.6%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds | Balanced accuracy | 70.9% (68.6%–73.0%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | sleep stage, five stages (SL-A) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +0.19 pp (−0.74–+1.14 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 54.4% (53.5%–55.4%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 58.9% (57.5%–60.2%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 57.5% (56.0%–58.8%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 53.2% (52.2%–54.3%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +3.02 pp (+1.55–+4.35 pp) — C1 higher; 2 pp margin not met (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | −1.41 pp (−2.40–−0.50 pp) — B-sh higher; B-sh non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-E)) | Paired difference, balanced accuracy | +1.20 pp (+0.07–+2.44 pp) — B-lin higher; B-lin non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds | Balanced accuracy | 63.5% (62.3%–64.7%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds | Balanced accuracy | 65.5% (64.1%–66.7%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds | Balanced accuracy | 65.2% (63.9%–66.5%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | next epoch’s stage differs (SL-E) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +1.75 pp (+0.92–+2.63 pp) — C1-fz higher; 2 pp margin not met; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 54.2% (53.0%–55.6%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 55.8% (54.2%–57.4%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 55.6% (54.2%–57.1%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 52.2% (51.0%–53.7%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +1.42 pp (−0.25–+2.90 pp) — No difference shown; 2 pp margin not met (not read: at floor): inconclusive at this sample size; at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | −0.14 pp (−1.41–+1.15 pp) — No difference shown; equivalent within ±2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-F)) | Paired difference, balanced accuracy | +2.01 pp (+0.54–+3.89 pp) — B-lin higher; B-lin non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · first or second half of the night (SL-F) · heads on frozen features, three seeds | Balanced accuracy | 64.5% (62.8%–66.1%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · first or second half of the night (SL-F) · heads on frozen features, three seeds | Balanced accuracy | 65.2% (63.6%–66.9%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · first or second half of the night (SL-F) · heads on frozen features, three seeds | Balanced accuracy | 65.2% (63.6%–66.9%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | first or second half of the night (SL-F) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +0.77 pp (−0.14–+1.67 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | awake or asleep (SL-B) · its own model, A(SL-B), against reading it off A(SL-A) | log R, the dedicated model’s remaining error over the read-out’s | −0.380 (−0.583–−0.203) — Dedicated model leaves less error; 20% margin not met; gate passed. | 100 | **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. SL-E and SL-F are largely predictable from the current stage: read off the true current stage alone, they reach 61.9% and 59.2% balanced accuracy (100 people; descriptive). A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ### Questions in language · sleep (route 3) Source page: [Questions in language](https://bci.report/topics/questions-in-language/) · [questions-in-language-update.json](https://bci.report/data/questions-in-language-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · label template − question number (TPL − ID) · seen, 5-way · three seeds | Paired difference, language arm minus reference (pp) | −0.05 pp (−0.70–+0.61 pp) — No difference shown; equivalent within ±2 pp; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · description − question number (DESC − ID) · seen, 5-way · three seeds | Paired difference, language arm minus reference (pp) | +0.02 pp (−0.59–+0.62 pp) — No difference shown; equivalent within ±2 pp; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new sentence frame − training wording (TPL(a1) − TPL(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −0.46 pp (−1.33–+0.16 pp) — No difference shown; equivalent within ±2 pp; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · synonym − training wording (TPL(a2) − TPL(train)) · seen, reworded, for these three synonyms · three seeds | Paired difference, language arm minus reference (pp) | −12.60 pp (−13.74–−11.43 pp) — TPL(train) higher; 2 pp margin not met; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new description − training descriptions (DESC(a3) − DESC(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −20.87 pp (−29.49–−13.71 pp) — DESC(train) higher; 2 pp margin not met; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · asleep · implicit name · three seeds | log R, the head’s remaining error over the read-off’s | +3.520 (+3.139–+3.967) — Unseen question. R 33.80; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · asleep · explicit name · three seeds | log R, the head’s remaining error over the read-off’s | +1.951 (+1.623–+2.357) — Unseen question. R 7.04; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · N1, N2 or N3 sleep · explicit name · three seeds | log R, the head’s remaining error over the read-off’s | +0.883 (+0.726–+1.047) — Unseen question. R 2.42; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · light sleep · implicit name · three seeds | log R, the head’s remaining error over the read-off’s | +2.267 (+2.072–+2.466) — Unseen question. R 9.65; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · light sleep · explicit name · three seeds | log R, the head’s remaining error over the read-off’s | +1.519 (+1.343–+1.698) — Unseen question. R 4.57; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · description against the read-off (DESC / read-off) · asleep · new descriptions · three seeds | log R, the head’s remaining error over the read-off’s | +3.444 (+2.913–+3.956) — Unseen question. R 31.30; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · description against the read-off (DESC / read-off) · N1, N2 or N3 sleep · new descriptions · three seeds | log R, the head’s remaining error over the read-off’s | +1.299 (+0.793–+1.744) — Unseen question. R 3.67; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · description against the read-off (DESC / read-off) · light sleep · new descriptions · three seeds | log R, the head’s remaining error over the read-off’s | +1.417 (+0.903–+1.856) — Unseen question. R 4.12; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P5 · label template − shuffled templates (TPL − SHUF) · unseen questions, implicit names · three seeds | Paired difference in AUROC | −0.378 (−0.501–−0.253) — Unseen question. SHUF higher; difference only: no margin; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P5 · label template − shuffled templates (TPL − SHUF) · unseen questions, explicit names · three seeds | Paired difference in AUROC | +0.629 (+0.603–+0.654) — Unseen question. TPL higher; difference only: no margin; headroom gate passed. | 100 | **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. **Limits:** 'Unseen' means unseen by the EEG head; the text encoder has read these words, and every token of a held-out frequency also occurs in the training texts. Held-out sleep questions are combinations of seen stages ('asleep' is the complement of one), reported as compositional beside the read-off. Rewording and unseen results are separate rows and never pooled. 35 pre-declared comparisons, not corrected for multiplicity. The pre-run permutation canary failed twice (revisions 3 and 4) and was revised by the owner, who overrode revision 4's "no further revision"; revision 5 passed. Both failed records are kept and disclosed. [Every limitation of route 3 →](https://bci.report/topics/questions-in-language/#methods-and-limits) ## Source and licence ### BOAS · Bitbrain Open Access Sleep dataset **Credit** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://openneuro.org/datasets/ds005555/versions/1.1.3) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. ### Ethics and consent, as the publisher states them The dataset description states approval by the Comité de Ética de la Investigación de la Comunidad Autónoma de Aragón (C.I. PI24/046). The README states that participants, adult members of the general population, provided written informed consent. The owner approved publishing BOAS cohort aggregates with stated gaps on 2026-10-07. Participants are pseudonymised in the public release. [Row in the open register →](https://bci.report/datasets/register/#r-site-boas) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *BOAS · Bitbrain Open Access Sleep dataset: EEG decoding results*. https://bci.report/datasets/boas/ Figures from releases [`questions-in-language-update-20261008`](https://bci.report/releases/#questions-in-language-update-20261008) (2026-10-08), [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007) (2026-10-07). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/boas/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Dreem Open Datasets (DOD-H and DOD-O): EEG decoding results Five-stage sleep staging, healthy sleepers and people with obstructive sleep apnoea, kept apart The Dreem Open Datasets are two overnight polysomnography cohorts released to compare human and automated sleep staging: DOD-H, 25 healthy volunteers recorded at the French Armed Forces Biomedical Research Institute (IRBA), and DOD-O, 55 patients with obstructive sleep apnoea recorded at the Stanford Sleep Medicine Center. Both were scored by five sleep technologists from different sleep centres. Described by Guillot, Sauvet, During and Thorey in IEEE Transactions on Neural Systems and Rehabilitation Engineering (2020), with the official dreem-learning-open code; archived on Zenodo (record 15900394, published July 2025). **Also known as** Dreem Open Datasets · DOD-H · DOD-O · DOD · Zenodo 15900394 · Guillot et al. (2020) Description sources [arxiv.org](https://arxiv.org/abs/1911.03221) · [doi.org](https://doi.org/10.1109/TNSRE.2020.3011181) · [zenodo.org](https://zenodo.org/records/15900394) · [github.com](https://github.com/Dreem-Organization/dreem-learning-open) ## Where it appears - [Why can a sleep stager be right most of the time and still miss whole stages?](https://bci.report/topics/sleep-staging/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Sleep staging · DOD-H · healthy sleepers (a separate experiment) Source page: [Sleep-stage balance](https://bci.report/topics/sleep-staging/) · [large-source-update.json](https://bci.report/data/large-source-update.json) **Balanced accuracy, 2 configurations.** Dot: the mean over nights; line: 95% interval. No dashed line: the training prior is a floor this baseline sets, not a chance level. - Training prior: 20.2% (20.0%–20.6%) - Spectral ridge: 56.7% (53.1%–60.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Training prior — comparator | DOD-H · healthy sleepers, every night held out once | Balanced accuracy | 20.2% (20.0%–20.6%) | 25 | | Training prior — comparator | DOD-H · healthy sleepers, every night held out once | Accuracy | 48.0% (44.6%–51.4%) | 25 | | Training prior — comparator | DOD-H · healthy sleepers, every night held out once | Macro F1 | 13.0% (12.3%–13.8%) | 25 | | Training prior — comparator | DOD-H · healthy sleepers, every night held out once | Cohen’s kappa | 0.000 (0.000–0.000) | 25 | | Spectral ridge | DOD-H · healthy sleepers, every night held out once | Balanced accuracy | 56.7% (53.1%–60.0%) | 25 | | Spectral ridge | DOD-H · healthy sleepers, every night held out once | Accuracy | 72.3% (67.8%–76.3%) | 25 | | Spectral ridge | DOD-H · healthy sleepers, every night held out once | Macro F1 | 53.3% (49.0%–57.3%) | 25 | | Spectral ridge | DOD-H · healthy sleepers, every night held out once | Cohen’s kappa | 0.584 (0.526–0.636) | 25 | | Spectral ridge | Spectral ridge minus training prior, the same nights | Paired difference, balanced accuracy | +36.5 pp (+32.9 pp–+39.8 pp) — 25 nights under both baselines; the interval excludes zero. | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage Wake | Recall | 64.8% (54.9%–73.9%) — 3,037 Wake epochs in the consensus. | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage Wake | Precision | 76.4% (70.1%–82.3%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage Wake | F1 | 66.5% (58.5%–73.8%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N1 | Recall | 0.0% (0.0%–0.0%) — 1,505 N1 epochs in the consensus; never predicted on any night, so every one is missed and N1 precision is not defined (not zero). | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N1 | F1 | 0.0% (0.0%–0.0%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N2 | Recall | 85.4% (77.4%–91.8%) — 11,879 N2 epochs in the consensus. | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N2 | Precision | 77.3% (72.8%–81.8%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N2 | F1 | 78.9% (74.0%–83.1%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N3 | Recall | 59.4% (46.8%–71.2%) — 3,514 N3 epochs in the consensus; defined on 24 of 25 nights. | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N3 | Precision | 69.7% (55.7%–81.9%) — defined on 24 of 25 nights. | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage N3 | F1 | 56.2% (43.6%–67.7%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage REM | Recall | 73.6% (63.8%–82.6%) — 4,727 REM epochs in the consensus. | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage REM | Precision | 63.8% (55.4%–71.5%) | 25 | | Spectral ridge | DOD-H · healthy sleepers · stage REM | F1 | 64.6% (56.1%–72.2%) | 25 | ### Sleep staging · DOD-O · people with obstructive sleep apnoea (a separate experiment) Source page: [Sleep-stage balance](https://bci.report/topics/sleep-staging/) · [large-source-update.json](https://bci.report/data/large-source-update.json) **Balanced accuracy, 2 configurations.** Dot: the mean over nights; line: 95% interval. No dashed line: the training prior is a floor this baseline sets, not a chance level. - Training prior: 20.5% (20.1%–21.2%) - Spectral ridge: 49.3% (47.3%–51.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Training prior — comparator | DOD-O · people with obstructive sleep apnoea, every night held out once | Balanced accuracy | 20.5% (20.1%–21.2%) | 55 | | Training prior — comparator | DOD-O · people with obstructive sleep apnoea, every night held out once | Accuracy | 49.2% (46.3%–52.2%) | 55 | | Training prior — comparator | DOD-O · people with obstructive sleep apnoea, every night held out once | Macro F1 | 13.3% (12.8%–13.9%) | 55 | | Training prior — comparator | DOD-O · people with obstructive sleep apnoea, every night held out once | Cohen’s kappa | 0.000 (0.000–0.000) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea, every night held out once | Balanced accuracy | 49.3% (47.3%–51.2%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea, every night held out once | Accuracy | 72.5% (69.7%–75.0%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea, every night held out once | Macro F1 | 47.4% (44.9%–49.7%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea, every night held out once | Cohen’s kappa | 0.542 (0.504–0.578) | 55 | | Spectral ridge | Spectral ridge minus training prior, the same nights | Paired difference, balanced accuracy | +28.8 pp (+26.7 pp–+30.7 pp) — 55 nights under both baselines; the interval excludes zero. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage Wake | Recall | 80.4% (75.5%–85.1%) — 10,427 Wake epochs in the consensus. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage Wake | Precision | 76.5% (72.0%–80.7%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage Wake | F1 | 75.6% (71.7%–79.3%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N1 | Recall | 0.0% (0.0%–0.0%) — 2,860 N1 epochs in the consensus; never predicted on any night, so every one is missed and N1 precision is not defined (not zero). | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N1 | F1 | 0.0% (0.0%–0.0%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N2 | Recall | 92.8% (90.5%–94.7%) — 26,271 N2 epochs in the consensus. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N2 | Precision | 71.0% (67.9%–74.0%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N2 | F1 | 79.8% (77.3%–82.0%) | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N3 | Recall | 22.1% (15.4%–29.2%) — 5,500 N3 epochs in the consensus; defined on 52 of 55 nights. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N3 | Precision | 63.4% (52.4%–73.9%) — defined on 49 of 55 nights. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage N3 | F1 | 27.9% (20.3%–35.8%) — defined on 53 of 55 nights. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage REM | Recall | 49.5% (43.1%–55.9%) — 8,103 REM epochs in the consensus; defined on 53 of 55 nights. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage REM | Precision | 68.2% (60.9%–74.9%) — defined on 54 of 55 nights. | 55 | | Spectral ridge | DOD-O · people with obstructive sleep apnoea · stage REM | F1 | 52.5% (46.0%–58.7%) — defined on 54 of 55 nights. | 55 | ## Source and licence ### Dreem Open Datasets (DOD-H and DOD-O) **Credit** Antoine Guillot, Fabien Sauvet, Emmanuel H. During and Valentin Thorey · Dreem Open Datasets: Multi-Scored Sleep Datasets to Compare Human and Automated Sleep Staging, IEEE Transactions on Neural Systems and Rehabilitation Engineering (2020), doi:10.1109/TNSRE.2020.3011181; preprint arXiv:1911.03221. Data: Dreem Open Datasets, Zenodo, doi:10.5281/zenodo.15900394. Official repository: github.com/Dreem-Organization/dreem-learning-open, revision 8b332d6827f5ae6a22f4bb97b4deef4273238ec3. [MIT ↗](https://opensource.org/licenses/MIT) · [Dataset record ↗](https://zenodo.org/records/15900394) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-dreem-dod) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Dreem Open Datasets (DOD-H and DOD-O): EEG decoding results*. https://bci.report/datasets/dreem-dod/ Figures from release [`large-source-update-20261003`](https://bci.report/releases/#large-source-update-20261003) (2026-10-03). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/dreem-dod/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ds003810: EEG decoding results Motor imagery / rest OpenNeuro ds003810 is an EEG dataset of kinesthetic motor imagery of dominant-hand grasping versus rest, recorded from 10 participants without previous BCI experience using a low-cost OpenBCI Cyton + Daisy board with 15 electrodes over sensorimotor areas. Released on OpenNeuro by Peterson et al. and described in Data in Brief (2022). **Also known as** ds003810 · Motor Imagery vs Rest - Low-Cost EEG System · A motor imagery vs. rest dataset with low-cost consumer grade EEG hardware · Peterson et al. (2022) Description sources [openneuro.org](https://openneuro.org/datasets/ds003810) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/) · [doi.org](https://doi.org/10.1016/j.dib.2022.108225) · [doi.org](https://doi.org/10.1016/j.heliyon.2020.e03425) ## Where it appears - [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) - [Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/) - [How often does an EEG decoder fire when nobody is giving a command?](https://bci.report/topics/when-not-to-act/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-mi-rest.csv](https://bci.report/data/core-matrix-20261106-mi-rest.csv) **Balanced accuracy, 9 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Spectral ridge: 54.9% (53.2%–56.7%) - CSP+LDA: 47.1% (43.8%–50.2%) - LaBraM: 53.5% (51.4%–55.4%) - CBraMod: 62.9% (59.6%–66.5%) - EEGNet: 71.0% (65.0%–76.3%) - TS+LR · OAS · 4–40 Hz: 58.9% (54.9%–63.9%) - MDM · OAS · 4–40 Hz: 53.8% (50.8%–57.9%) - ShallowFBCSPNet: 68.8% (63.4%–73.9%) - Deep4Net: 68.6% (63.3%–74.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Rest versus right-hand imagery | Balanced accuracy | 54.9% (53.2%–56.7%) | 10 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Rest versus right-hand imagery | Balanced accuracy | 47.1% (43.8%–50.2%) | 10 | | [LaBraM](https://bci.report/methods/labram/) | Rest versus right-hand imagery | Balanced accuracy | 53.5% (51.4%–55.4%) | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | Rest versus right-hand imagery | Balanced accuracy | 62.9% (59.6%–66.5%) | 10 | | [EEGNet](https://bci.report/methods/eegnet/) | Rest versus right-hand imagery | Balanced accuracy | 71.0% (65.0%–76.3%) | 10 | | [TS+LR · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Rest versus right-hand imagery | Balanced accuracy | 58.9% (54.9%–63.9%) — Primary row · one fixed configuration (MOABB-reference settings) | 10 | | [MDM · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Rest versus right-hand imagery | Balanced accuracy | 53.8% (50.8%–57.9%) — Secondary row · one fixed configuration (MOABB-reference settings) | 10 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Rest versus right-hand imagery | Balanced accuracy | 68.8% (63.4%–73.9%) — One fixed seed and training budget, as the EEGNet cell | 10 | | [Deep4Net](https://bci.report/methods/deep4net/) | Rest versus right-hand imagery | Balanced accuracy | 68.6% (63.3%–74.0%) — One fixed seed and training budget, as the EEGNet cell | 10 | ### Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) **Balanced accuracy, 16 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 66.2% (61.8%–70.6%) - REVE Large · Frozen encoder + ridge head: 70.2% (64.0%–76.1%) - LUNA Base · Frozen encoder + ridge head: 52.6% (50.2%–54.9%) - LUNA Large · Frozen encoder + ridge head: 49.9% (46.6%–53.5%) - BrainOmni Base · Frozen encoder + ridge head: 58.8% (57.0%–60.8%) - CodeBrain · Frozen encoder + ridge head: 58.8% (54.5%–62.6%) - EEGMamba · Frozen encoder + ridge head: 67.3% (63.6%–70.6%) - ST-EEGFormer Base · Frozen encoder + ridge head: 58.1% (53.7%–63.3%) - ST-EEGFormer Large · Frozen encoder + ridge head: 62.4% (57.4%–67.8%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 68.1% (64.3%–72.3%) - ERP-FM Base · Frozen encoder + ridge head: 54.8% (51.8%–57.8%) - SingLEM · Frozen encoder + ridge head: 72.0% (66.6%–76.6%) - ZUNA 1.1 · Frozen encoder + ridge head: 50.7% (49.2%–52.1%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 63.2% (58.2%–67.9%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 64.9% (60.3%–69.4%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 63.3% (59.0%–67.4%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 66.2% (61.8%–70.6%) | 10 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 70.2% (64.0%–76.1%) | 10 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 52.6% (50.2%–54.9%) | 10 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 49.9% (46.6%–53.5%) | 10 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 58.8% (57.0%–60.8%) | 10 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 58.8% (54.5%–62.6%) | 10 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 67.3% (63.6%–70.6%) | 10 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 58.1% (53.7%–63.3%) | 10 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 62.4% (57.4%–67.8%) | 10 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 68.1% (64.3%–72.3%) | 10 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 54.8% (51.8%–57.8%) | 10 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 72.0% (66.6%–76.6%) | 10 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 50.7% (49.2%–52.1%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 10 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 63.2% (58.2%–67.9%) | 10 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 64.9% (60.3%–69.4%) | 10 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 63.3% (59.0%–67.4%) | 10 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### EEGPT, frozen encoder · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 58.3% (54.9%–61.8%) | 10 | ds003810 is recorded at 125 Hz; its windows reach the encoder upsampled to 256 Hz, so they carry nothing above 62.5 Hz. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/mi-rest/#eegpt-terms) ### Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json), [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · pretrained | Balanced accuracy | 53.5% | 10 | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · random initialization (mean) | Balanced accuracy | 55.9% | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · pretrained | Balanced accuracy | 62.9% | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · random initialization (mean) | Balanced accuracy | 56.5% | 10 | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · pretrained | Balanced accuracy | 58.3% | 10 | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · random initialization (mean) | Balanced accuracy | 54.1% | 10 | ### Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json), [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · pretrained | Balanced accuracy | 52.9% | 10 | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · random initialization (mean) | Balanced accuracy | 56.5% | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · pretrained | Balanced accuracy | 62.7% | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · random initialization (mean) | Balanced accuracy | 54.9% | 10 | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · pretrained | Balanced accuracy | 57.0% | 10 | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · random initialization (mean) | Balanced accuracy | 53.1% | 10 | ### Pretraining · fixed classical control Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Log-covariance ridge | MI / rest | Balanced accuracy | 52.3% (48.5%–56.4%) | 10 | ### Reliable decisions · when to decline a decision (crude: ten people) Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 0 / 1,200 — 10 of 10 people had nothing accepted. | 10 | | Spectral ridge | Learned reject option minus calibrated confidence, same classifier (a secondary contrast) | Error among the 80% most certain, difference | +0.5 pp (−1.1 pp–+2.3 pp) — No difference resolved: the interval includes zero. | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 7 / 1,200 — 6 of 10 people had nothing accepted. | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | Learned reject option minus calibrated confidence, same classifier (a secondary contrast) | Error among the 80% most certain, difference | +0.7 pp (−0.8 pp–+2.4 pp) — No difference resolved: the interval includes zero. | 10 | | [EEGNet](https://bci.report/methods/eegnet/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 250 / 1,200 — 0 of 10 people had nothing accepted. | 10 | | [EEGNet](https://bci.report/methods/eegnet/) | Learned reject option minus calibrated confidence, same classifier (a secondary contrast) | Error among the 80% most certain, difference | +0.9 pp (−0.4 pp–+2.4 pp) — No difference resolved: the interval includes zero. | 10 | **Limits:** Ten people, two per test fold. Bootstrap intervals are crude; no person-level dispersion is published. Balanced, so no deployment rate. Its learned-reject contrasts are secondary. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ## Source and licence ### ds003810 **Credit** Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ **The dataset authors ask you to cite:** Peterson V, Galván C, Hernández H, Spies R. A feasibility study of a complete low-cost consumer-grade brain-computer interface system. Heliyon 6(3):e03425 (2020). [doi ↗](https://doi.org/10.1016/j.heliyon.2020.e03425) · [as their record asks ↗](https://openneuro.org/datasets/ds003810) [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ds003810) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *ds003810: EEG decoding results*. https://bci.report/datasets/ds003810/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ds003810/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # OpenNeuro ds004584 · Parkinson's disease, rest eyes open: EEG decoding results Resting-state EEG, Parkinson's disease and controls Two minutes of eyes-open resting-state EEG (64-channel BrainVision cap) from 149 participants, 100 with Parkinson's disease and 49 controls, recruited at the University of Iowa Movement Disorders Clinic (2017 to 2021) for the cohort of Singh et al. (J Neurol Neurosurg Psychiatry, 2023). Released on OpenNeuro as ds004584 'Rest eyes open' (CC0, 2023). **Also known as** OpenNeuro ds004584 · Rest eyes open · NEMAR on004584 Description sources [openneuro.org](https://openneuro.org/datasets/ds004584) · [doi.org](https://doi.org/10.18112/openneuro.ds004584.v1.0.0) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC10592174/) · [doi.org](https://doi.org/10.1101/2022.07.26.22278079) · [nemar.org](https://nemar.org/dataset/on004584) ## Where it appears - [Can resting-state EEG separate Parkinson's disease from controls?](https://bci.report/topics/clinical-groups/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Clinical groups · Parkinson's disease vs. controls Source page: [Clinical groups](https://bci.report/topics/clinical-groups/) · Chance level 50.0% · [clinical-update.json](https://bci.report/data/clinical-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Relative-spectrum logistic regression · Resting-state EEG, eyes open: 71.7% (63.5%–79.3%) - Age and sex only · Age and sex only · no EEG: 55.0% (46.9%–63.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Relative-spectrum logistic regression | Resting-state EEG, eyes open | Balanced accuracy | 71.7% (63.5%–79.3%) | 149 | | Relative-spectrum logistic regression | Resting-state EEG, eyes open | AUROC | 0.76 | 149 | | Age and sex only — comparator | Age and sex only · no EEG | Balanced accuracy | 55.0% (46.9%–63.6%) | 149 | | Age and sex only — comparator | Age and sex only · no EEG | AUROC | 0.56 | 149 | ## Source and licence ### OpenNeuro ds004584 · Parkinson's disease, rest eyes open **Credit** Dataset: Arun Singh, Rachel Cole, Arturo Espinoza, Jim Cavanagh and Nandakumar Narayanan · Rest eyes open, OpenNeuro ds004584 v1.0.0, doi:10.18112/openneuro.ds004584.v1.0.0. Cohort study: Arun Singh, Rachel C. Cole, Arturo I. Espinoza, Jan R. Wessel, James F. Cavanagh and Nandakumar S. Narayanan · Evoked mid-frontal activity predicts cognitive dysfunction in Parkinson's disease, Journal of Neurology, Neurosurgery & Psychiatry 94, 945-953 (2023), doi:10.1136/jnnp-2022-330154. Transport mirror: NEMAR on004584, doi:10.82901/nemar.on004584. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds004584.v1.0.0) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ds004584) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *OpenNeuro ds004584 · Parkinson's disease, rest eyes open: EEG decoding results*. https://bci.report/datasets/ds004584/ Figures from release [`clinical-update-20260923`](https://bci.report/releases/#clinical-update-20260923) (2026-09-23). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ds004584/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ds005342: EEG decoding results Cue-gated idle / command OpenNeuro ds005342 contains 17-channel EEG (250 Hz) from 32 healthy participants who, following on-screen cues, imagined sit-to-stand or stand-to-sit movements or stayed in an idle state, first in offline calibration runs and then in online BCI runs. Released on OpenNeuro by Triana-Guzman et al. (2024); the study appeared in Frontiers in Neuroinformatics (2022). **Also known as** ds005342 · EEG data offline and online during motor imagery for standing and sitting · Triana-Guzman et al. (2022) Description sources [openneuro.org](https://openneuro.org/datasets/ds005342) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC9481272/) · [doi.org](https://doi.org/10.3389/fninf.2022.961089) ## Where it appears - [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-idle.csv](https://bci.report/data/core-matrix-20261106-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 23.3% | 4 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Seated motor imagery · cue-gated replay | Idle false activation | 0.0% | 4 | | [LaBraM](https://bci.report/methods/labram/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 18.3% | 4 | | [LaBraM](https://bci.report/methods/labram/) | Seated motor imagery · cue-gated replay | Idle false activation | 3.3% | 4 | | [EEGNet](https://bci.report/methods/eegnet/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 1.7% | 4 | | [EEGNet](https://bci.report/methods/eegnet/) | Seated motor imagery · cue-gated replay | Idle false activation | 0.0% | 4 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 43.3% | 4 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% | 4 | | [CBraMod](https://bci.report/methods/cbramod/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 18.3% | 4 | | [CBraMod](https://bci.report/methods/cbramod/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% | 4 | | [Deep4Net](https://bci.report/methods/deep4net/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 3.3% | 4 | | [Deep4Net](https://bci.report/methods/deep4net/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% | 4 | | [TS+LR · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 35.0% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | | [TS+LR · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Idle false activation | 6.7% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | | [MDM · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 35.0% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | | [MDM · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | ### Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + linear head | Commands detected ≤3 s | 12 / 60 — 0 of 4 people always abstained. | 4 | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + linear head | Commands detected ≤3 s | 6 / 60 — 2 of 4 people always abstained. | 4 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + linear head | Commands detected ≤3 s | 4 / 60 — 2 of 4 people always abstained. | 4 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + linear head | Commands detected ≤3 s | 5 / 60 — 2 of 4 people always abstained. | 4 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + linear head | Commands detected ≤3 s | 13 / 60 — 1 of 4 people always abstained. | 4 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + linear head | Idle false activations | 3 / 60 | 4 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + linear head | Commands detected ≤3 s | 10 / 60 — 1 of 4 people always abstained. | 4 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + linear head | Commands detected ≤3 s | 2 / 60 — 3 of 4 people always abstained. | 4 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + linear head | Idle false activations | 0 / 60 | 4 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Commands detected ≤3 s | 16 / 60 — 1 of 4 people always abstained. | 4 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Commands detected ≤3 s | 20 / 60 — 1 of 4 people always abstained. | 4 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head | Commands detected ≤3 s | 3 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head | Idle false activations | 0 / 60 | 4 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + linear head | Commands detected ≤3 s | 3 / 60 — 2 of 4 people always abstained. | 4 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + linear head | Commands detected ≤3 s | 6 / 60 — 2 of 4 people always abstained. | 4 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + linear head | Idle false activations | 5 / 60 | 4 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + linear head | Commands detected ≤3 s | 11 / 60 — 0 of 4 people always abstained. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 4 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + linear head | Idle false activations | 1 / 60 — Model card: research use only, not for diagnosis or clinical use. | 4 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Commands detected ≤3 s | 12 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Idle false activations | 0 / 60 | 4 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Commands detected ≤3 s | 5 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Idle false activations | 0 / 60 | 4 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Commands detected ≤3 s | 5 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Idle false activations | 0 / 60 | 4 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) ### EEGPT, frozen encoder · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#eegpt) · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Commands detected ≤3 s | 7 / 60 — 1 of 4 people always abstained. | 4 | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Idle false activations | 2 / 60 | 4 | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Window balanced accuracy | 57.2% | 4 | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Window AUROC | 0.780 | 4 | Same run as the published LaBraM idle row; the published CBraMod idle row comes from another run, so a comparison with it is across runs. The window balanced accuracy and AUROC were computed from this run’s saved probabilities with the definitions of the further foundation-encoder rows; the run itself recorded the counts. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/idle/#eegpt-terms) ## Source and licence ### ds005342 **Credit** OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ds005342) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *ds005342: EEG decoding results*. https://bci.report/datasets/ds005342/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ds005342/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ds006593: EEG decoding results P300 target ERP The cBCI Matrix Multimodal Dataset (OpenNeuro ds006593) contains EEG and eye-tracking calibration recordings from 21 control participants performing a single-character-presentation matrix speller task in BciPy, recorded at Northeastern University in 2023 with a DSI-24 dry-electrode cap. Released on OpenNeuro in 2025 by Celik et al. **Also known as** ds006593 · cBCI Matrix Multimodal Dataset · Multimodal Sensor Fusion for EEG-Based BCI Typing Systems · on006593 Description sources [openneuro.org](https://openneuro.org/datasets/ds006593) · [nemar.org](https://nemar.org/dataset/on006593) ## Where it appears - [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) **Balanced accuracy, 10 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Spectral ridge: 51.0% (49.6%–52.4%) - Temporal ridge: 53.8% (51.9%–55.6%) - LaBraM: 49.4% (47.7%–51.2%) - CBraMod: 55.0% (53.3%–56.7%) - EEGNet: 53.3% (51.7%–54.9%) - XDAWNCov+TS+LR · 4 filters · 50 Hz: 61.9% (58.8%–65.0%) - xDAWN+LDA · 4 filters · 50 Hz: 60.4% (57.7%–63.2%) - ShallowFBCSPNet: 49.7% (49.0%–50.4%) - Deep4Net: 50.8% (50.0%–51.6%) - CSP+LDA: 50.0% (49.1%–50.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Visual target versus nontarget events | Balanced accuracy | 51.0% (49.6%–52.4%) | 21 | | Temporal ridge | Visual target versus nontarget events | Balanced accuracy | 53.8% (51.9%–55.6%) | 21 | | [LaBraM](https://bci.report/methods/labram/) | Visual target versus nontarget events | Balanced accuracy | 49.4% (47.7%–51.2%) | 21 | | [CBraMod](https://bci.report/methods/cbramod/) | Visual target versus nontarget events | Balanced accuracy | 55.0% (53.3%–56.7%) | 21 | | [EEGNet](https://bci.report/methods/eegnet/) | Visual target versus nontarget events | Balanced accuracy | 53.3% (51.7%–54.9%) | 21 | | [XDAWNCov+TS+LR · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Visual target versus nontarget events | Balanced accuracy | 61.9% (58.8%–65.0%) — Primary row · one fixed configuration (MOABB-reference settings) | 21 | | [xDAWN+LDA · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Visual target versus nontarget events | Balanced accuracy | 60.4% (57.7%–63.2%) — Secondary row · one fixed configuration (MOABB-reference settings) | 21 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Visual target versus nontarget events | Balanced accuracy | 49.7% (49.0%–50.4%) — One fixed seed and training budget, as the EEGNet cell | 21 | | [Deep4Net](https://bci.report/methods/deep4net/) | Visual target versus nontarget events | Balanced accuracy | 50.8% (50.0%–51.6%) — Braindecode's automatically scaled Deep4Net for 1-s input (no temporal pooling, 229-sample classifier, 235,752 parameters), not the original Deep4Net configuration. The interval's lower end, 50.04 %, prints as 50.0; it sits 0.04 pp above the 50 % reference. | 21 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Visual target versus nontarget events | Balanced accuracy | 50.0% (49.1%–50.8%) — Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. On this ERP protocol it reads band power, not the waveform. | 21 | ### Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) **Balanced accuracy, 15 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 50.5% (48.5%–52.3%) - REVE Large · Frozen encoder + ridge head: 54.6% (52.7%–56.5%) - LUNA Base · Frozen encoder + ridge head: 48.7% (46.3%–51.1%) - LUNA Large · Frozen encoder + ridge head: 48.8% (46.9%–50.8%) - CodeBrain · Frozen encoder + ridge head: 50.8% (49.3%–52.2%) - EEGMamba · Frozen encoder + ridge head: 52.5% (51.4%–53.5%) - ST-EEGFormer Base · Frozen encoder + ridge head: 53.6% (52.3%–54.8%) - ST-EEGFormer Large · Frozen encoder + ridge head: 52.7% (51.3%–54.2%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 49.5% (48.1%–51.0%) - ERP-FM Base · Frozen encoder + ridge head: 51.2% (49.3%–53.2%) - SingLEM · Frozen encoder + ridge head: 51.2% (49.2%–53.1%) - ZUNA 1.1 · Frozen encoder + ridge head: 51.6% (50.3%–53.0%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 51.6% (50.4%–52.7%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 51.1% (49.2%–53.0%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 50.0% (48.7%–51.5%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 50.5% (48.5%–52.3%) | 21 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 54.6% (52.7%–56.5%) | 21 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 48.7% (46.3%–51.1%) | 21 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 48.8% (46.9%–50.8%) | 21 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 50.8% (49.3%–52.2%) | 21 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 52.5% (51.4%–53.5%) | 21 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 53.6% (52.3%–54.8%) | 21 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 52.7% (51.3%–54.2%) | 21 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 49.5% (48.1%–51.0%) | 21 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 51.2% (49.3%–53.2%) | 21 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 51.2% (49.2%–53.1%) | 21 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 51.6% (50.3%–53.0%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 21 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 51.6% (50.4%–52.7%) | 21 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 51.1% (49.2%–53.0%) | 21 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 50.0% (48.7%–51.5%) | 21 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; CodeBrain and ZUNA under Apache-2.0; EEGMamba, ST-EEGFormer and SingLEM under MIT; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) **BrainOmni Base: not run.** BrainTokenizer needs a 512-sample window (2.0 s at 256 Hz); the published segment is 1 s = 256 samples, so upstream unfold() would zero-pad 256 of 512 samples of every window (invented samples). All upstream downstream tasks use >= 2 s windows. ### EEGPT, frozen encoder · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 54.0% (51.5%–56.6%) | 21 | On the P300 and semantic targets, one-second windows ran as four real 0.25-second patches: the patch grid was set to four after the encoder was built, with no padding. EEGPT’s encoder treats each time patch on its own, so this is numerically the same as building it for 256-sample inputs, as the Orthosis pilot built it for its 128-sample windows. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/p300-target/#eegpt-terms) ## Source and licence ### ds006593 **Credit** OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ds006593) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *ds006593: EEG decoding results*. https://bci.report/datasets/ds006593/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ds006593/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # OpenNeuro ds007420 · multi-distance fNIRS: fNIRS results Human fNIRS recordings: a numerical coverage check, no decoding OpenNeuro ds007420 holds multi-distance, high-density fNIRS over the motor cortex from 12 healthy adults during a resting-state task, ball squeezing with both hands, and a motion-artifact task of six instructed movements (head tilts, raised eyebrows and saying “hello”), with accelerometer recordings. Released in BIDS format by Shibo Zhou, De’Ja Rogers, David A. Boas, Meryem A. Yücel and Yuanyuan Gao under CC0; the dataset asks users to acknowledge Gao et al., Neurophotonics (2023). Only a label-independent numerical check of its optical values is run here. **Also known as** ds007420 · Open-access Multi-Distance fNIRS Dataset Generated During Ball-Squeezing and Motion-Artifact Induction Tasks · Zhou et al. multi-distance fNIRS Description sources [doi.org](https://doi.org/10.18112/openneuro.ds007420.v1.0.5) · [raw.githubusercontent.com](https://raw.githubusercontent.com/OpenNeuroDatasets/ds007420/1.0.5/dataset_description.json) · [europepmc.org](https://europepmc.org/article/PMC/PMC10203730) ## Where it appears - [Do simple baselines work on signals other than scalp EEG?](https://bci.report/topics/beyond-scalp-eeg/#ds007420) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Beyond scalp EEG · human fNIRS recordings, numerical coverage counts Source page: [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/#ds007420) · [beyond-scalp-update.json](https://bci.report/data/beyond-scalp-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Numerical coverage check (no model) | Every optical value of every recording | Finite and positive values | 48,502,600 / 48,502,600 — Across 60 recordings and 12,000 recording-channels; 0 non-finite, 0 zero and 0 negative values. | 12 | | Numerical coverage check (no model) | Every channel of every recording | Exactly constant channels | 0 / 12,000 | 12 | Counts only, with no measure of signal quality, neural origin or decoding. No classifier score exists: the task codes have no verified lookup to hands or movements, and the event durations disagree — 1.0 s in the events table, 10.0 s in the optical files, 5.0 s in the cited paper. ## Source and licence ### OpenNeuro ds007420 · multi-distance fNIRS **Credit** Shibo Zhou, De’Ja Rogers, David A. Boas, Meryem A. Yücel and Yuanyuan Gao · Open-access Multi-Distance fNIRS Dataset Generated During Ball-Squeezing and Motion-Artifact Induction Tasks, OpenNeuro, doi:10.18112/openneuro.ds007420.v1.0.5, CC0. The dataset asks to acknowledge: Gao, Y., Rogers, D., von Lühmann, A., Ortega-Martinez, A., Boas, D.A., & Yücel, M.A. (2023). Short-separation regression incorporated diffuse optical tomography image reconstruction modeling for high-density functional near-infrared spectroscopy. Neurophotonics, 10(2), 025007. https://doi.org/10.1117/1.NPh.10.2.025007. Derived analysis by BCI Report; not endorsed by the authors. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds007420.v1.0.5) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ds007420) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *OpenNeuro ds007420 · multi-distance fNIRS: fNIRS results*. https://bci.report/datasets/ds007420/ Figures from release [`beyond-scalp-update-20261012`](https://bci.report/releases/#beyond-scalp-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ds007420/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEG-fNIRS-based Handwriting Trajectory Dataset: EEG decoding results Four released coded labels in an imagined-handwriting competition dataset, EEG only: a data-quality and methods case study The EEG-fNIRS-based Handwriting Trajectory Dataset is the data release of an imagined-handwriting competition, posted on Hugging Face by the account lasfk: raw EEG and fNIRS recordings with trial tables, twenty labelled training participants and ten test participants, and four class labels mapped from released event codes. The card names no organisers and no paper; a participating team's challenge paper identifies the competition as the ACM MM 2026 Multimodal Brain-Computer Interface Grand Challenge. The card declares CC BY-NC-SA 4.0, behind a gate. Only the EEG of five training participants is used here, and the official test data are not. **Also known as** lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset · ACM MM 2026 Multimodal Brain-Computer Interface Grand Challenge Description sources [huggingface.co](https://huggingface.co/datasets/lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset) · [arxiv.org](https://arxiv.org/abs/2608.03176v2) ## Where it appears - [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/#handwriting) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Pilot studies · mask-aware EEG on an imagined-handwriting dataset, a data-quality and methods case study Source page: [Pilot studies](https://bci.report/topics/pilot-studies/#handwriting) · Chance level 25.00% · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Mask-aware relative EEG + logistic regression | A person left out of training · four released coded labels | Balanced accuracy | 26.81% (26.30%–27.20%) | 5 | | Missingness mask only (no EEG) — comparator | A person left out of training · four released coded labels | Balanced accuracy | 24.99% — Interval not published: with five people, its bounds would show individual people’s scores. | 5 | | Elapsed time / trial order (no EEG) — comparator | A person left out of training · four released coded labels | Balanced accuracy | 25.44% (24.03%–26.79%) | 5 | | Mask + time (no EEG) — comparator | A person left out of training · four released coded labels | Balanced accuracy | 25.01% (23.90%–26.12%) | 5 | | Training majority — comparator | A person left out of training · four released coded labels | Balanced accuracy | 25.00% — Chance for every person, by construction. | 5 | | Mask-aware relative EEG + logistic regression | Minus the missingness mask only, the same people | Paired difference | +1.82 pp (+1.30–+2.21 pp) | 5 | | Mask-aware relative EEG + logistic regression | Minus elapsed time / trial order, the same people | Paired difference | +1.37 pp (−0.12–+3.07 pp) — The interval crosses zero: not separated from this control. | 5 | | Mask-aware relative EEG + logistic regression | Minus mask + time, the same people | Paired difference | +1.80 pp (+0.43–+3.05 pp) | 5 | | Training-label shuffles · Mask-aware relative EEG + logistic regression — comparator | A person left out of training · four released coded labels | Balanced accuracy | 25.20% — The mean of 5 shuffles, whose means ranged from 23.85% to 26.69%: an observed range, not a confidence interval. | 5 | 5 people and 1,618 trials from a reused development cohort, whose earlier scores were seen before this design: a quality-informed diagnostic, not independent confirmation. Every trial keeps 27 to 38 of 59 recorded channels under a fixed numerical availability rule, which is not certified clean EEG. Four released coded labels, not verified characters or trajectories: not handwriting recognition. The intervals are descriptive, resampling the five people over overlapping folds, not population confidence intervals. The official test data were not used, and the physiological origin of the signal is not established. ## Source and licence ### EEG-fNIRS-based Handwriting Trajectory Dataset **Credit** lasfk on Hugging Face · EEG-fNIRS-based Handwriting Trajectory Dataset, revision 4036680b5fb0015472d55589da2f34c5a87a3c10, CC BY-NC-SA 4.0 as declared on the dataset card, gated access. Derived analysis by BCI Report; not endorsed by the uploader. [CC BY-NC-SA 4.0 ↗](https://creativecommons.org/licenses/by-nc-sa/4.0/) · [Dataset record ↗](https://huggingface.co/datasets/lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. ### Ethics and consent No ethics-approval or informed-consent statement was found, on the dataset card or in a participating team’s challenge paper; the aggregates are published with that gap stated, by the owner’s decision. [The full statement on the data-use page →](https://bci.report/data-use/#handwriting-rights) [Row in the open register →](https://bci.report/datasets/register/#r-site-eeg-fnirs-handwriting) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *EEG-fNIRS-based Handwriting Trajectory Dataset: EEG decoding results*. https://bci.report/datasets/eeg-fnirs-handwriting/ Figures from release [`pilot-studies-update-20261012`](https://bci.report/releases/#pilot-studies-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/eeg-fnirs-handwriting/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEGMAT: EEG decoding results Mental arithmetic / rest EEGMAT (EEG During Mental Arithmetic Tasks) is a PhysioNet database, version 1.0.0 published December 2018, with two 60-second artifact-free EEG recordings from each of 36 participants: one before and one during a serial-subtraction task. Contributed from Igor Sikorsky Kyiv Polytechnic Institute and described by Zyma et al. in Data (2019). **Also known as** EEGMAT · EEG During Mental Arithmetic Tasks · Electroencephalograms during Mental Arithmetic Task Performance · Zyma et al. (2019) Description sources [physionet.org](https://physionet.org/content/eegmat/1.0.0/) · [doi.org](https://doi.org/10.3390/data4010014) ## Where it appears - [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) - [Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/) - [New people, next day: which part of a pretrained model should you update?](https://bci.report/topics/model-adaptation/) - [How often does an EEG decoder fire when nobody is giving a command?](https://bci.report/topics/when-not-to-act/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) **Balanced accuracy, 9 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Spectral ridge: 56.8% (54.6%–58.9%) - LaBraM: 64.6% (60.5%–68.6%) - CBraMod: 62.3% (58.1%–66.5%) - EEGNet: 67.6% (62.8%–72.5%) - TS+LR · OAS · 4–40 Hz: 65.5% (59.6%–71.8%) - MDM · OAS · 4–40 Hz: 71.5% (65.5%–77.8%) - ShallowFBCSPNet: 67.9% (63.4%–72.4%) - Deep4Net: 61.3% (57.5%–65.2%) - CSP+LDA: 68.3% (62.0%–74.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Serial subtraction versus resting EEG | Balanced accuracy | 56.8% (54.6%–58.9%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | Serial subtraction versus resting EEG | Balanced accuracy | 64.6% (60.5%–68.6%) | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Serial subtraction versus resting EEG | Balanced accuracy | 62.3% (58.1%–66.5%) | 36 | | [EEGNet](https://bci.report/methods/eegnet/) | Serial subtraction versus resting EEG | Balanced accuracy | 67.6% (62.8%–72.5%) | 36 | | [TS+LR · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Serial subtraction versus resting EEG | Balanced accuracy | 65.5% (59.6%–71.8%) — Primary row · one fixed configuration (MOABB-reference settings) | 36 | | [MDM · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Serial subtraction versus resting EEG | Balanced accuracy | 71.5% (65.5%–77.8%) — Secondary row · one fixed configuration (MOABB-reference settings) | 36 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Serial subtraction versus resting EEG | Balanced accuracy | 67.9% (63.4%–72.4%) — One fixed seed and training budget, as the EEGNet cell | 36 | | [Deep4Net](https://bci.report/methods/deep4net/) | Serial subtraction versus resting EEG | Balanced accuracy | 61.3% (57.5%–65.2%) — One fixed seed and training budget, as the EEGNet cell | 36 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Serial subtraction versus resting EEG | Balanced accuracy | 68.3% (62.0%–74.3%) — Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. | 36 | ### Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) **Balanced accuracy, 16 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 68.8% (64.1%–73.2%) - REVE Large · Frozen encoder + ridge head: 69.0% (63.3%–74.7%) - LUNA Base · Frozen encoder + ridge head: 60.7% (56.9%–64.4%) - LUNA Large · Frozen encoder + ridge head: 57.4% (54.3%–60.6%) - BrainOmni Base · Frozen encoder + ridge head: 61.6% (58.4%–64.9%) - CodeBrain · Frozen encoder + ridge head: 63.0% (59.5%–66.5%) - EEGMamba · Frozen encoder + ridge head: 60.3% (57.0%–63.4%) - ST-EEGFormer Base · Frozen encoder + ridge head: 64.5% (59.0%–70.0%) - ST-EEGFormer Large · Frozen encoder + ridge head: 63.4% (57.8%–69.0%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 63.6% (59.1%–68.1%) - ERP-FM Base · Frozen encoder + ridge head: 61.1% (57.3%–65.1%) - SingLEM · Frozen encoder + ridge head: 51.5% (49.2%–53.8%) - ZUNA 1.1 · Frozen encoder + ridge head: 53.3% (49.5%–57.1%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 61.6% (56.4%–66.5%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 64.3% (60.2%–68.2%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 62.5% (58.5%–66.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 68.8% (64.1%–73.2%) | 36 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 69.0% (63.3%–74.7%) | 36 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 60.7% (56.9%–64.4%) | 36 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 57.4% (54.3%–60.6%) | 36 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 61.6% (58.4%–64.9%) | 36 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 63.0% (59.5%–66.5%) | 36 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 60.3% (57.0%–63.4%) | 36 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 64.5% (59.0%–70.0%) | 36 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 63.4% (57.8%–69.0%) | 36 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 63.6% (59.1%–68.1%) | 36 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 61.1% (57.3%–65.1%) | 36 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 51.5% (49.2%–53.8%) | 36 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 53.3% (49.5%–57.1%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 61.6% (56.4%–66.5%) | 36 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 64.3% (60.2%–68.2%) | 36 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 62.5% (58.5%–66.2%) | 36 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### EEGPT, frozen encoder · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 65.7% (60.9%–70.3%) | 36 | Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/arithmetic-rest/#eegpt-terms) ### Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json), [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · pretrained | Balanced accuracy | 64.6% | 36 | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · random initialization (mean) | Balanced accuracy | 56.5% | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · pretrained | Balanced accuracy | 62.3% | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · random initialization (mean) | Balanced accuracy | 56.4% | 36 | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · pretrained | Balanced accuracy | 65.7% | 36 | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · random initialization (mean) | Balanced accuracy | 59.2% | 36 | ### Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json), [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · pretrained | Balanced accuracy | 64.4% | 36 | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · random initialization (mean) | Balanced accuracy | 56.5% | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · pretrained | Balanced accuracy | 68.4% | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · random initialization (mean) | Balanced accuracy | 61.7% | 36 | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · pretrained | Balanced accuracy | 66.0% | 36 | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · random initialization (mean) | Balanced accuracy | 58.6% | 36 | ### Pretraining · fixed classical control Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Log-covariance ridge | Mental workload | Balanced accuracy | 61.9% (56.2%–67.6%) | 36 | ### Model adaptation · LaBraM on new people, same task, zero labels from the test person Source page: [Which part to update?](https://bci.report/topics/model-adaptation/) · Chance level 50.0% · [adaptation-update.json](https://bci.report/data/adaptation-update.json) **Balanced accuracy, 3 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LaBraM · Head only · encoder frozen: 56.6% (54.2%–59.0%) - LaBraM · Last block + head: 65.7% (61.6%–69.7%) - LaBraM · LoRA rank 4 + head: 64.1% (60.2%–67.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Head only · encoder frozen | Balanced accuracy | 56.6% (54.2%–59.0%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | Last block + head | Balanced accuracy | 65.7% (61.6%–69.7%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | LoRA rank 4 + head | Balanced accuracy | 64.1% (60.2%–67.8%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | LoRA rank 4 + head minus head only | Paired difference | +7.5 pp (+4.8 pp–+10.1 pp) | 36 | | [LaBraM](https://bci.report/methods/labram/) | Last block + head minus head only | Paired difference | +9.1 pp (+6.0 pp–+12.2 pp) | 36 | | [LaBraM](https://bci.report/methods/labram/) | LoRA rank 4 + head minus last block + head | Paired difference | −1.6 pp (−4.4 pp–+1.0 pp) | 36 | ### Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + trained head | Balanced accuracy | 64.0% (60.4%–67.7%) — 1,026 trainable parameters. | 36 | | [REVE Base](https://bci.report/methods/reve/) | LoRA rank 4 + head | Balanced accuracy | 66.3% (61.9%–70.7%) — 181,250 trainable parameters. | 36 | | [REVE Base](https://bci.report/methods/reve/) | LoRA minus frozen + head, same people and folds | Paired difference | +2.3 pp (+0.1 pp–+4.6 pp) — 21 people improved, 13 got worse, 2 unchanged. The interval excludes zero. | 36 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + trained head | Balanced accuracy | 65.6% (61.1%–70.1%) — 2,434 trainable parameters. | 36 | | [REVE Large](https://bci.report/methods/reve/) | LoRA rank 4 + head | Balanced accuracy | 69.2% (63.4%–74.9%) — 430,466 trainable parameters. | 36 | | [REVE Large](https://bci.report/methods/reve/) | LoRA minus frozen + head, same people and folds | Paired difference | +3.6 pp (+0.5 pp–+6.8 pp) — 21 people improved, 13 got worse, 2 unchanged. The interval excludes zero. | 36 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + trained head | Balanced accuracy | 58.2% (54.5%–61.7%) — 514 trainable parameters. | 36 | | [LUNA Base](https://bci.report/methods/luna/) | LoRA rank 4 + head | Balanced accuracy | 60.0% (55.9%–63.9%) — 37,378 trainable parameters. | 36 | | [LUNA Base](https://bci.report/methods/luna/) | LoRA minus frozen + head, same people and folds | Paired difference | +1.8 pp (+0.1 pp–+3.3 pp) — 23 people improved, 11 got worse, 2 unchanged. The interval excludes zero. | 36 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + trained head | Balanced accuracy | 64.5% (59.6%–69.2%) — 16,386 trainable parameters. | 36 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | LoRA rank 4 + head | Balanced accuracy | 64.8% (59.6%–69.7%) — 106,498 trainable parameters. | 36 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | LoRA minus frozen + head, same people and folds | Paired difference | +0.3 pp (−1.4 pp–+1.8 pp) — 18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established. | 36 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + trained head | Balanced accuracy | 66.5% (62.3%–70.6%) — 7,602 trainable parameters. | 36 | | [CodeBrain](https://bci.report/methods/codebrain/) | LoRA rank 4 + head | Balanced accuracy | 67.1% (62.6%–71.6%) — 58,802 trainable parameters. | 36 | | [CodeBrain](https://bci.report/methods/codebrain/) | LoRA minus frozen + head, same people and folds | Paired difference | +0.6 pp (−0.7 pp–+1.9 pp) — 18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established. | 36 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + trained head | Balanced accuracy | 60.4% (55.7%–65.2%) — 7,602 trainable parameters. | 36 | | [EEGMamba](https://bci.report/methods/eegmamba/) | LoRA rank 4 + head | Balanced accuracy | 60.8% (56.3%–65.3%) — 90,930 trainable parameters. | 36 | | [EEGMamba](https://bci.report/methods/eegmamba/) | LoRA minus frozen + head, same people and folds | Paired difference | +0.4 pp (−0.7 pp–+1.5 pp) — 18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established. | 36 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + trained head | Balanced accuracy | 63.2% (59.7%–66.7%) — 1,538 trainable parameters. | 36 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | LoRA rank 4 + head | Balanced accuracy | 65.1% (61.0%–69.4%) — 148,994 trainable parameters. | 36 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | LoRA minus frozen + head, same people and folds | Paired difference | +1.9 pp (−0.2 pp–+4.1 pp) — 20 people improved, 16 got worse, 0 unchanged. The interval includes zero: no change is established. | 36 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + trained head | Balanced accuracy | 50.9% (49.5%–52.3%) — 1,218 trainable parameters. | 36 | | [SingLEM](https://bci.report/methods/singlem/) | LoRA rank 4 + head | Balanced accuracy | 60.6% (56.7%–64.4%) — 33,986 trainable parameters. | 36 | | [SingLEM](https://bci.report/methods/singlem/) | LoRA minus frozen + head, same people and folds | Paired difference | +9.7 pp (+6.1 pp–+13.3 pp) — 30 people improved, 6 got worse, 0 unchanged. The interval excludes zero. | 36 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + trained head | Balanced accuracy | 51.2% (49.8%–52.7%) — 1,218 trainable parameters. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | LoRA rank 4 + head | Balanced accuracy | 56.9% (53.7%–60.1%) — 296,130 trainable parameters. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | LoRA minus frozen + head, same people and folds | Paired difference | +5.7 pp (+2.5 pp–+8.7 pp) — 28 people improved, 7 got worse, 1 unchanged. The interval excludes zero. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 3 / 2,160 — 33 of 36 people had nothing accepted. | 36 | | Spectral ridge | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.5 pp (−1.2 pp–+1.7 pp) — No difference resolved: the interval includes zero. | 36 | | [EEGNet](https://bci.report/methods/eegnet/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 126 / 2,160 — 19 of 36 people had nothing accepted. | 36 | | [EEGNet](https://bci.report/methods/eegnet/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +2.3 pp (+0.6 pp–+4.1 pp) — Difference resolved: the interval excludes zero. | 36 | | [LaBraM · head only, encoder frozen](https://bci.report/methods/labram/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 1 / 2,160 — 35 of 36 people had nothing accepted. | 36 | | [LaBraM · head only, encoder frozen](https://bci.report/methods/labram/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.6 pp (−0.5 pp–+2.7 pp) — No difference resolved: the interval includes zero. | 36 | | [LaBraM · LoRA rank 4 + head](https://bci.report/methods/labram/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 44 / 2,160 — 24 of 36 people had nothing accepted. | 36 | | [LaBraM · LoRA rank 4 + head](https://bci.report/methods/labram/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +1.9 pp (+0.6 pp–+3.5 pp) — Difference resolved: the interval excludes zero. | 36 | **Limits:** Balanced by design. Error rates describe a 1:1 mix and are a method comparison, not a deployment rate. Class equals recording, so recording-level differences are part of what any model or selector can use. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ## Source and licence ### EEGMAT **Credit** Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) · [Dataset record ↗](https://physionet.org/content/eegmat/1.0.0/) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-eegmat) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *EEGMAT: EEG decoding results*. https://bci.report/datasets/eegmat/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`adaptation-update-20261001`](https://bci.report/releases/#adaptation-update-20261001) (2026-10-01), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/eegmat/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EESM19 scalp subset: EEG decoding results Five-stage sleep EESM19 (Ear-EEG Sleep Monitoring 2019, OpenNeuro ds005185) is a home sleep dataset from Aarhus University: 20 participants each slept four nights with dry-contact ear-EEG plus partial polysomnography, manually sleep-scored, and 10 of them added 12 ear-EEG-only nights. First published by Mikkelsen et al. in Scientific Reports (2019). **Also known as** EESM19 · Ear-EEG Sleep Monitoring 2019 · ds005185 · Mikkelsen et al. (2019) Description sources [openneuro.org](https://openneuro.org/datasets/ds005185) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC6856384/) · [doi.org](https://doi.org/10.1038/s41598-019-53115-3) ## Where it appears - [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) - [Can one EEG model answer several questions about the same data as well as separate models, and at what cost?](https://bci.report/topics/shared-encoder/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-sleep-scalp.csv](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) **Balanced accuracy, 7 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Spectral ridge: 72.2% (70.7%–73.8%) - LaBraM: 71.1% (69.0%–73.2%) - CBraMod: 71.3% (68.7%–73.7%) - EEGNet: 56.5% (54.9%–58.3%) - ShallowFBCSPNet: 55.0% (53.1%–56.8%) - Deep4Net: 45.2% (42.7%–47.7%) - CSP+LDA (one-vs-rest CSP): 50.1% (46.9%–53.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Five stages · balanced scalp-EEG sample | Balanced accuracy | 72.2% (70.7%–73.8%) | 20 | | [LaBraM](https://bci.report/methods/labram/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 71.1% (69.0%–73.2%) | 20 | | [CBraMod](https://bci.report/methods/cbramod/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 71.3% (68.7%–73.7%) | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 56.5% (54.9%–58.3%) | 20 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 55.0% (53.1%–56.8%) — One fixed seed and training budget, as the EEGNet cell | 20 | | [Deep4Net](https://bci.report/methods/deep4net/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 45.2% (42.7%–47.7%) — One fixed seed and training budget, as the EEGNet cell | 20 | | [CSP+LDA (one-vs-rest CSP)](https://bci.report/methods/csp-lda/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 50.1% (46.9%–53.1%) — One-vs-rest CSP: one binary CSP per stage, 30 log-variance features, one shrinkage LDA; broadband as in the published mi-rest recipe; a new code path, not a tuned CSP. | 20 | ### Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) **Balanced accuracy, 16 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 74.6% (72.9%–76.4%) - REVE Large · Frozen encoder + ridge head: 76.6% (75.2%–78.0%) - LUNA Base · Frozen encoder + ridge head: 67.5% (65.3%–69.5%) - LUNA Large · Frozen encoder + ridge head: 72.4% (70.4%–74.0%) - BrainOmni Base · Frozen encoder + ridge head: 67.7% (65.8%–69.5%) - CodeBrain · Frozen encoder + ridge head: 70.7% (68.8%–72.7%) - EEGMamba · Frozen encoder + ridge head: 64.1% (61.6%–66.4%) - ST-EEGFormer Base · Frozen encoder + ridge head: 77.3% (75.6%–78.8%) - ST-EEGFormer Large · Frozen encoder + ridge head: 79.6% (78.2%–81.0%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 67.9% (65.4%–70.4%) - ERP-FM Base · Frozen encoder + ridge head: 71.0% (69.5%–72.6%) - SingLEM · Frozen encoder + ridge head: 35.3% (33.7%–37.0%) - ZUNA 1.1 · Frozen encoder + ridge head: 60.1% (57.8%–62.3%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 70.3% (68.0%–72.7%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 65.2% (63.0%–67.3%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 68.3% (65.7%–70.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 74.6% (72.9%–76.4%) | 20 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 76.6% (75.2%–78.0%) | 20 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 67.5% (65.3%–69.5%) | 20 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 72.4% (70.4%–74.0%) | 20 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 67.7% (65.8%–69.5%) | 20 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 70.7% (68.8%–72.7%) | 20 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 64.1% (61.6%–66.4%) | 20 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 77.3% (75.6%–78.8%) | 20 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 79.6% (78.2%–81.0%) | 20 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 67.9% (65.4%–70.4%) | 20 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 71.0% (69.5%–72.6%) | 20 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 35.3% (33.7%–37.0%) | 20 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 60.1% (57.8%–62.3%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 20 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 70.3% (68.0%–72.7%) | 20 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 65.2% (63.0%–67.3%) | 20 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 68.3% (65.7%–70.8%) | 20 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; BrainOmni, EEGMamba, ST-EEGFormer and SingLEM under MIT; CodeBrain and ZUNA under Apache-2.0; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### EEGPT, frozen encoder · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#eegpt) · Chance level 20.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 75.5% (73.9%–77.1%) | 20 | Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/sleep-scalp/#eegpt-terms) ### One model, several questions · sleep, the full EESM19 release (crude: 20 people, one seed) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · sleep stage, five stages (SL-A) · trained from scratch, one seed | Balanced accuracy | 63.2% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · sleep stage, five stages (SL-A) · trained from scratch, one seed | Balanced accuracy | 66.7% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · sleep stage, five stages (SL-A) · trained from scratch, one seed | Balanced accuracy | 64.7% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +3.47 pp (+1.54–+5.64 pp) — C1 higher; 2 pp margin not met; gate passed. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-A)) | Paired difference, balanced accuracy | −1.47 pp (−2.90–−0.10 pp) — A(SL-A) higher; 2 pp margin not met; gate passed. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · next epoch’s stage differs (SL-E) · trained from scratch, one seed | Balanced accuracy | 53.4% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · next epoch’s stage differs (SL-E) · trained from scratch, one seed | Balanced accuracy | 56.8% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · next epoch’s stage differs (SL-E) · trained from scratch, one seed | Balanced accuracy | 52.1% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +3.38 pp (+2.27–+4.36 pp) — C1 higher; 2 pp margin not met (not read: at floor); at floor: reported, not counted. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-E)) | Paired difference, balanced accuracy | +1.33 pp (+0.42–+2.32 pp) — B-lin higher; B-lin non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · first or second half of the night (SL-F) · trained from scratch, one seed | Balanced accuracy | 61.8% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · first or second half of the night (SL-F) · trained from scratch, one seed | Balanced accuracy | 63.8% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · first or second half of the night (SL-F) · trained from scratch, one seed | Balanced accuracy | 62.5% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +2.09 pp (+0.38–+3.74 pp) — C1 higher; 2 pp margin not met; gate passed. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-F)) | Paired difference, balanced accuracy | −0.72 pp (−2.10–+0.69 pp) — No difference shown; 2 pp margin not met: inconclusive at this sample size; gate passed. | 20 | **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). EESM19 full (S13) uses every stored epoch that passes the published core rules at its natural stage mix; it is a different protocol from the site's balanced sleep-scalp subset. SL-E and SL-F are largely predictable from the current stage; no stage-only read-out was computed on EESM19. A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ## Source and licence ### EESM19 scalp subset **Credit** Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-eesm19) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *EESM19 scalp subset: EEG decoding results*. https://bci.report/datasets/eesm19/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007) (2026-10-07), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/eesm19/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EESM23 · Ear-EEG Sleep Monitoring 2023: EEG decoding results Five-stage sleep, in-ear and scalp Ear-EEG Sleep Monitoring 2023 (EESM23): at-home sleep recordings from 10 healthy participants by an Aarhus University group, each with two nights of combined ear-EEG and polysomnography followed by ten ear-EEG-only nights, using generic ear pieces with dry electrodes. Released on OpenNeuro as ds005178 (CC0) and described in Scientific Data (2025). **Also known as** EESM23 · Ear-EEG Sleep Monitoring 2023 · OpenNeuro ds005178 Description sources [openneuro.org](https://openneuro.org/datasets/ds005178) · [doi.org](https://doi.org/10.18112/openneuro.ds005178.v1.0.0) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC11840015/) ## Where it appears - [Can fewer electrodes, or electrodes in the ear, match a full scalp montage?](https://bci.report/topics/fewer-electrodes/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Fewer electrodes · in-ear vs. scalp sleep staging Source page: [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) · Chance level 20.0% · [evidence-update.json](https://bci.report/data/evidence-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Log-bandpower logistic regression · Four in-ear channels: 53.6% (47.4%–59.7%) - Log-bandpower logistic regression · Six scalp electrodes: 69.0% (63.4%–74.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Log-bandpower logistic regression | Four in-ear channels | Balanced accuracy | 53.6% (47.4%–59.7%) | 10 | | Log-bandpower logistic regression | Six scalp electrodes | Balanced accuracy | 69.0% (63.4%–74.2%) | 10 | | Log-bandpower logistic regression | Scalp minus in-ear, same epochs | Paired difference | +15.4 pp (+10.6 pp–+20.3 pp) | 10 | ## Source and licence ### EESM23 · Ear-EEG Sleep Monitoring 2023 **Credit** Dataset: Yousef Rezaei Tabar, Kaare Mikkelsen, Laura Birch, Nelly Shenton, Simon L. Kappel, Astrid R. Bertelsen, Reza Nikbakht, Hans O. Toft, Chris H. Henriksen, Martin C. Hemmsen, Mike L. Rank, Marit Otto and Preben Kidmose · Ear-EEG Sleep Monitoring 2023 (EESM23), OpenNeuro ds005178 v1.0.0, doi:10.18112/openneuro.ds005178.v1.0.0. Study: Kaare Bjarke Mikkelsen, Yousef Rezai Tabar, Laura Rævsbæk Birch, Simon Lind Kappel, Christian Bech Christensen, Lars Dalskov Mosgaard, Marit Otto, Martin Christian Hemmsen, Mike Lind Rank and Preben Kidmose · Ear-EEG sleep monitoring data sets, Scientific Data 12, 301 (2025), doi:10.1038/s41597-025-04579-8. Processed export: Zachary1150/EESM23-Processed on Hugging Face, which did not create the original cohort. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005178.v1.0.0) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-eesm23) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *EESM23 · Ear-EEG Sleep Monitoring 2023: EEG decoding results*. https://bci.report/datasets/eesm23/ Figures from release [`evidence-update-20260922`](https://bci.report/releases/#evidence-update-20260922) (2026-09-22). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/eesm23/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Forenzo & He continuous-tracking EEG-BCI dataset: EEG decoding results Continuous cursor tracking with a noninvasive BCI, across sessions The Forenzo & He continuous-tracking dataset (KiltHub 25360300, Carnegie Mellon University) holds EEG from 28 participants who used motor imagery to steer a cursor after a target moving continuously across a screen (continuous pursuit), over several sessions with traditional and deep-learning decoders, in two substudies, the second testing transfer learning; recorded with a 64-channel cap at 1 kHz. Released by Dylan Forenzo and Bin He under CC BY 4.0 and described in PNAS Nexus (2024). Only offline analyses from each record's earliest to its latest complete session are run here; the paper's online results are not reproduced. **Also known as** Forenzo & He (2024) · KiltHub 25360300 · EEG-BCI Dataset for Continuous Tracking using Deep Learning-based Decoding for Non-invasive Brain-Computer Interface · Forenzo et al. (2024) · continuous pursuit BCI dataset Description sources [doi.org](https://doi.org/10.1184/R1/25360300.v1) · [doi.org](https://doi.org/10.1093/pnasnexus/pgae145) · [api.figshare.com](https://api.figshare.com/v2/articles/25360300/versions/1) ## Where it appears - [Does a decoder trained on an earlier session still work later?](https://bci.report/topics/later-sessions/#forenzo) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Later sessions · continuous cursor tracking, Main cohort, historical decoder velocity (imitates the historical decoder, not intended motion; a negative result, offline) Source page: [Later sessions](https://bci.report/topics/later-sessions/#forenzo) · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Historical decoder velocity · earliest → latest session | Normalized RMSE (lower is better) | 192.290 (1.292–574.235) — Median 1.317: the mean is far above it, a severe upper tail. | 9 records | | Source-mean comparator | Historical decoder velocity · earliest → latest session | Normalized RMSE (lower is better) | 1.266 (1.214–1.316) | 9 records | | Spectral ridge | Historical decoder velocity · ridge minus comparator, the same records | Paired difference in normalized RMSE | 191.024 (0.026–572.995) — The ridge had the larger error on 9 of 9 admitted records; 9 of 14 candidate records admitted, the rest held on metadata before scoring, so the result is conditional on that subset. | 9 records | The People column counts records, which are not proven to be unique people. ### Later sessions · continuous cursor tracking, Main cohort, constructed displacement proxy (a negative result, offline) Source page: [Later sessions](https://bci.report/topics/later-sessions/#forenzo) · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Constructed displacement proxy · earliest → latest session | Normalized RMSE (lower is better) | 698.499 (1.166–2,093) — Median 1.225: the mean is far above it, a severe upper tail. | 9 records | | Source-mean comparator | Constructed displacement proxy · earliest → latest session | Normalized RMSE (lower is better) | 1.056 (0.977–1.133) | 9 records | | Spectral ridge | Constructed displacement proxy · ridge minus comparator, the same records | Paired difference in normalized RMSE | 697.443 (0.121–2,092) — The ridge had the larger error on 9 of 9 admitted records; 9 of 14 candidate records admitted, the rest held on metadata before scoring, so the result is conditional on that subset. | 9 records | ### Later sessions · continuous cursor tracking, Transfer Learning cohort, historical decoder velocity (imitates the historical decoder, not intended motion; a negative result, offline) Source page: [Later sessions](https://bci.report/topics/later-sessions/#forenzo) · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Historical decoder velocity · earliest → latest session | Normalized RMSE (lower is better) | 44.172 (0.955–130.476) — Median 0.990: the mean is far above it, a severe upper tail. | 14 records | | Source-mean comparator | Historical decoder velocity · earliest → latest session | Normalized RMSE (lower is better) | 0.864 (0.801–0.929) | 14 records | | Spectral ridge | Historical decoder velocity · ridge minus comparator, the same records | Paired difference in normalized RMSE | 43.308 (0.105–129.601) — The ridge had the larger error on 14 of 14 admitted records; 14 of 14 candidate records admitted. “Transfer Learning” is the publisher’s name for how the data were collected, not a model trained here. | 14 records | ### Later sessions · continuous cursor tracking, Transfer Learning cohort, constructed displacement proxy (a negative result, offline) Source page: [Later sessions](https://bci.report/topics/later-sessions/#forenzo) · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Constructed displacement proxy · earliest → latest session | Normalized RMSE (lower is better) | 69.844 (1.071–207.287) — Median 1.182: the mean is far above it, a severe upper tail. | 14 records | | Source-mean comparator | Constructed displacement proxy · earliest → latest session | Normalized RMSE (lower is better) | 0.916 (0.841–0.987) | 14 records | | Spectral ridge | Constructed displacement proxy · ridge minus comparator, the same records | Paired difference in normalized RMSE | 68.928 (0.143–206.410) — The ridge had the larger error on 14 of 14 admitted records; 14 of 14 candidate records admitted. “Transfer Learning” is the publisher’s name for how the data were collected, not a model trained here. | 14 records | ## Source and licence ### Forenzo & He continuous-tracking EEG-BCI dataset **Credit** Dylan Forenzo and Bin He · EEG-BCI Dataset for “Continuous Tracking using Deep Learning-based Decoding for Non-invasive Brain-Computer Interface”, KiltHub, Carnegie Mellon University, doi:10.1184/R1/25360300.v1, CC BY 4.0. Citation the record requests: Dylan Forenzo, Hao Zhu, Jenn Shanahan, Jaehyun Lim and Bin He, Continuous tracking using deep learning-based decoding for noninvasive brain–computer interface, PNAS Nexus 3(4), pgae145 (2024), doi:10.1093/pnasnexus/pgae145. Derived analysis by BCI Report; not endorsed by the authors and not a reproduction of the paper's online results. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.1184/R1/25360300.v1) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-forenzo-continuous-tracking) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Forenzo & He continuous-tracking EEG-BCI dataset: EEG decoding results*. https://bci.report/datasets/forenzo-continuous-tracking/ Figures from release [`later-sessions-update-20261008`](https://bci.report/releases/#later-sessions-update-20261008) (2026-10-08). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/forenzo-continuous-tracking/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Multimodal gait dataset · treadmill walking: EEG decoding results Treadmill walking at three speeds Treadmill-walking dataset from 59 healthy young adults at Khalifa University, each walking one minute at 0.5, 0.75 and 1.0 m/s, with synchronized 19-channel dry-electrode EEG (DSI-24), surface EMG from 12 lower-limb muscles, 17-sensor IMU kinematics and split-belt force plates. Released on PhysioNet as v1.0.0 (CC BY 4.0, April 2026) by Katmah et al. **Also known as** A multimodal gait dataset of brain activity, muscle activity, kinematics and ground forces in young adults · PhysioNet multimodal-gait-dataset · doi:10.13026/r0ea-7161 Description sources [physionet.org](https://physionet.org/content/multimodal-gait-dataset/1.0.0/) · [doi.org](https://doi.org/10.13026/r0ea-7161) ## Where it appears - [Does EEG decoding still work while walking or running?](https://bci.report/topics/on-the-move/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Movement · treadmill walking speed (a confound case) Source page: [On the move](https://bci.report/topics/on-the-move/) · Chance level 33.3% · [context-update.json](https://bci.report/data/context-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Relative spectral bands · Relative spectral bands, 19 scalp channels: 45.4% (39.1%–51.7%) - Movement-nuisance features · Movement-nuisance features, not brain signal: 50.6% (44.3%–56.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Relative spectral bands | Relative spectral bands, 19 scalp channels | Balanced accuracy | 45.4% (39.1%–51.7%) | 58 | | Movement-nuisance features — comparator | Movement-nuisance features, not brain signal | Balanced accuracy | 50.6% (44.3%–56.9%) | 58 | ## Source and licence ### Multimodal gait dataset · treadmill walking **Credit** R. Katmah, A. AlShehhi, D. Kosaji, N. Al-Rahmani, M. Abdullah, A. A. V. Hulleck and K. Khalaf · A multimodal gait dataset of brain activity, muscle activity, kinematics and ground forces in young adults, PhysioNet v1.0.0 (2026), doi:10.13026/r0ea-7161. [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.13026/r0ea-7161) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-gait-eeg) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Multimodal gait dataset · treadmill walking: EEG decoding results*. https://bci.report/datasets/gait-eeg/ Figures from release [`context-update-20260927`](https://bci.report/releases/#context-update-20260927) (2026-09-27). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/gait-eeg/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Kim beta-range SSVEP · 40 targets, 14.0–21.8 Hz (Figshare 28806815): EEG decoding results 40-target SSVEP speller, 14.0 to 21.8 Hz Kim beta-range SSVEP (Figshare 28806815) is the release of a 40-target SSVEP speller study with its stimuli in the beta range, described by Heegyu Kim, Kyungho Won, Minkyu Ahn and Sung Chan Jun in Scientific Data (2025) and deposited by Heegyu Kim under CC BY 4.0. Besides the task recordings, the record holds rest and questionnaire files; only the 40 task files are used here. **Also known as** Kim2025BetaRange · Figshare 28806815 · doi:10.1038/s41597-025-06032-2 - Forty-target SSVEP speller selection with stimuli from 14.0 to 21.8 Hz, six blocks per person on one day. - Cohort: 40 participants, 240 blocks, 9,600 trials. - All 40 task files passed their MD5 checks; the rest and questionnaire files were not acquired. The files store the targets in the order 14 + j + 0.2·i Hz, which differs from the ascending list in MOABB’s metadata, and the references follow the stored order. - Figshare record 28806815, version 3, the latest of the 3 versions the record lists. Description sources [doi.org](https://doi.org/10.6084/m9.figshare.28806815.v3) · [www.nature.com](https://www.nature.com/articles/s41597-025-06032-2) ## Where it appears - [What do the files show once a public EEG dataset is checked end to end?](https://bci.report/topics/data-quality/#kim2025-beta-range) The training-free CCA baseline, with its intake, the checks run before any score, its caveats and its hold, is the worked case on the data-quality page: [What does training-free CCA give on 40 frequencies in the beta range? →](https://bci.report/topics/data-quality/#kim2025-beta-range) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Data quality · which of 40 targets, with no training Source page: [Checked end to end](https://bci.report/topics/data-quality/#kim2025-beta-range) · Chance level 2.50% · [dataset-checks-update.json](https://bci.report/data/dataset-checks-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | CCA (no training) | 40 targets, 14.0–21.8 Hz · a nominal 2-second window from the stored onset, 11 channels | Accuracy | 53.40% (45.89%–61.01%) | 40 | ## Source and licence ### Kim beta-range SSVEP · 40 targets, 14.0–21.8 Hz (Figshare 28806815) **Credit** Heegyu Kim · SSVEP 40class dataset, figshare, doi:10.6084/m9.figshare.28806815.v3, CC BY 4.0. Paper: Heegyu Kim, Kyungho Won, Minkyu Ahn and Sung Chan Jun, A 40-Class SSVEP Speller Dataset: Beta Range Stimulation for Low-Fatigue BCI Applications, Scientific Data 12:1751 (2025), doi:10.1038/s41597-025-06032-2. Modifications: the 40 task files only; the mean of M1 and M2 subtracted, 11 channels, a 2-second window and a band-pass filter inside it; no data redistributed. Derived analysis by BCI Report; not endorsed by the authors. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.6084/m9.figshare.28806815.v3) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. ### Ethics and consent, as the source states them **Ethics approval:** “The Institutional Review Board at Gwangju Institute of Science and Technology (GIST) approved this experiment (20211201-HR-64-02-04)” **Informed consent:** “all subjects were informed of all experimental procedures and signed consent forms” The data descriptor also states: “All research data has been consented to by participants, with explicit agreement for public sharing.” The confidentiality clause of the consent forms excludes questionnaire-related personal data; the research acquired the 40 task files only and excluded the questionnaire and rest files. The paper’s Data availability links version 2; the research used version 3 (the latest), which adds a questionnaire. Both versions carry CC BY 4.0 in the licence field. [Data descriptor, with its ethics statement ↗](https://www.nature.com/articles/s41597-025-06032-2) [Row in the open register →](https://bci.report/datasets/register/#r-site-kim2025-beta-range) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Kim beta-range SSVEP · 40 targets, 14.0–21.8 Hz (Figshare 28806815): EEG decoding results*. https://bci.report/datasets/kim2025-beta-range/ Figures from release [`dataset-checks-update-20261028`](https://bci.report/releases/#dataset-checks-update-20261028) (2026-10-28). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/kim2025-beta-range/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Longitudinal ERP dataset (RSVP): EEG decoding results RSVP face-target ERP, the same people at nominal Day 1, 7, 80 and 200 visits A longitudinal EEG dataset of event-related potential (figshare 27201003) holds EEG from a rapid serial visual presentation task in which participants identified a target face: 15 people (Group A) recorded at four visits labelled Day 1, 7, 80 and 200, and 52 people (Group B) recorded once, with 57 EEG channels. Released by Chen Yang, Yufeng Zhang and colleagues (Beijing University of Posts and Telecommunications) under CC0 and described in Scientific Data (2025). Only Group A is used here. **Also known as** A longitudinal EEG dataset of event-related potential · figshare 27201003 · Longitudinal ERP dataset · Zhang et al. (2025) Description sources [doi.org](https://doi.org/10.6084/m9.figshare.27201003.v1) · [doi.org](https://doi.org/10.1038/s41597-025-05378-x) · [api.figshare.com](https://api.figshare.com/v2/articles/27201003/versions/1) ## Where it appears - [Does a decoder trained on an earlier session still work later?](https://bci.report/topics/later-sessions/#longitudinal-rsvp) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Later sessions · RSVP target detection, first visit to later visits Source page: [Later sessions](https://bci.report/topics/later-sessions/#longitudinal-rsvp) · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Normalized ERP features + logistic/Platt (CPU baseline) | Trained at the first visit · nominal Day 7 | AUROC | 0.888 (0.864–0.911) — Targets: 1,794 of 72,000 events. | 15 | | Normalized ERP features + logistic/Platt (CPU baseline) | Trained at the first visit · nominal Day 7 | Average precision | 0.359 | 15 | | Normalized ERP features + logistic/Platt (CPU baseline) | Trained at the first visit · nominal Day 80 | AUROC | 0.879 (0.854–0.903) — Targets: 1,786 of 72,000 events. | 15 | | Normalized ERP features + logistic/Platt (CPU baseline) | Trained at the first visit · nominal Day 80 | Average precision | 0.310 | 15 | | Normalized ERP features + logistic/Platt (CPU baseline) | Trained at the first visit · nominal Day 200 | AUROC | 0.853 (0.827–0.879) — Targets: 1,793 of 72,000 events. | 15 | | Normalized ERP features + logistic/Platt (CPU baseline) | Trained at the first visit · nominal Day 200 | Average precision | 0.240 | 15 | | Normalized ERP features + logistic/Platt (CPU baseline) | Day 200 minus Day 7, the same people | AUROC difference | −0.035 (−0.065–−0.006) — 6 of 15 people lost 0.05 AUROC or more. Nominal visit labels: no claim that elapsed time caused it. | 15 | ## Source and licence ### Longitudinal ERP dataset (RSVP) **Credit** Chen Yang, Yufeng Zhang, Hongxin Zhang, Yixuan Li, Yijun Wang and Xiaorong Gao · A longitudinal EEG dataset of event-related potential, figshare, doi:10.6084/m9.figshare.27201003.v1, CC0. Paper: Yufeng Zhang, Hongxin Zhang, Yixuan Li, Yijun Wang, Xiaorong Gao and Chen Yang, A longitudinal EEG dataset of event-related potential, Scientific Data 12, 1069 (2025), doi:10.1038/s41597-025-05378-x. Derived analysis by BCI Report; not endorsed by the authors. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.6084/m9.figshare.27201003.v1) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-longitudinal-rsvp) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Longitudinal ERP dataset (RSVP): EEG decoding results*. https://bci.report/datasets/longitudinal-rsvp/ Figures from release [`later-sessions-update-20261008`](https://bci.report/releases/#later-sessions-update-20261008) (2026-10-08). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/longitudinal-rsvp/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # LTRSVP · EEG Signals from an RSVP Task: EEG decoding results P300 target images in rapid serial visual presentation at three rates EEG Signals from an RSVP Task (PhysioNet ltrsvp, version 1.0.0, May 2017) holds EEG from 11 healthy participants who watched aerial images of London in rapid serial visual presentation at 5, 6 and 10 images per second; target images had an airplane superimposed. Eight posterior channels (PO8, PO7, PO3, PO4, P7, P8, O1, O2) at 2048 Hz, filtered 0.15–28 Hz, two recordings per rate separated by a long break. Contributed from the University of Essex and described by Matran-Fernandez and Poli in PLoS ONE (2017). **Also known as** LTRSVP · EEG Signals from an RSVP Task · PhysioNet ltrsvp · Matran-Fernandez & Poli (2017) Description sources [physionet.org](https://physionet.org/content/ltrsvp/1.0.0/) · [doi.org](https://doi.org/10.1371/journal.pone.0178498) ## Where it appears - [Does a P300 decoder calibrated on a screen still work in VR?](https://bci.report/topics/screen-to-vr/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Image rate · trained on one recording at one rate, tested on a different recording Source page: [Screen to VR](https://bci.report/topics/screen-to-vr/) · Chance level 50.0% · [extension-update.json](https://bci.report/data/extension-update.json) **Balanced accuracy, 9 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Logistic regression on 100-ms means · Trained at 5 Hz (run a), tested at 5 Hz (run b): 73.6% (65.8%–80.8%) - Logistic regression on 100-ms means · Trained at 5 Hz (run a), tested at 6 Hz (run b): 70.7% (64.5%–76.7%) - Logistic regression on 100-ms means · Trained at 5 Hz (run a), tested at 10 Hz (run b): 60.8% (56.5%–65.8%) - Logistic regression on 100-ms means · Trained at 6 Hz (run a), tested at 5 Hz (run b): 74.9% (68.5%–80.9%) - Logistic regression on 100-ms means · Trained at 6 Hz (run a), tested at 6 Hz (run b): 71.4% (64.3%–78.1%) - Logistic regression on 100-ms means · Trained at 6 Hz (run a), tested at 10 Hz (run b): 60.6% (57.4%–64.3%) - Logistic regression on 100-ms means · Trained at 10 Hz (run a), tested at 5 Hz (run b): 67.5% (59.9%–75.2%) - Logistic regression on 100-ms means · Trained at 10 Hz (run a), tested at 6 Hz (run b): 69.8% (61.9%–77.2%) - Logistic regression on 100-ms means · Trained at 10 Hz (run a), tested at 10 Hz (run b): 63.0% (57.7%–68.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Logistic regression on 100-ms means | Trained at 5 Hz (run a), tested at 5 Hz (run b) | Balanced accuracy | 73.6% (65.8%–80.8%) | 9 | | Logistic regression on 100-ms means | Trained at 5 Hz (run a), tested at 6 Hz (run b) | Balanced accuracy | 70.7% (64.5%–76.7%) | 9 | | Logistic regression on 100-ms means | Trained at 5 Hz (run a), tested at 10 Hz (run b) | Balanced accuracy | 60.8% (56.5%–65.8%) | 9 | | Logistic regression on 100-ms means | Trained at 6 Hz (run a), tested at 5 Hz (run b) | Balanced accuracy | 74.9% (68.5%–80.9%) | 9 | | Logistic regression on 100-ms means | Trained at 6 Hz (run a), tested at 6 Hz (run b) | Balanced accuracy | 71.4% (64.3%–78.1%) | 9 | | Logistic regression on 100-ms means | Trained at 6 Hz (run a), tested at 10 Hz (run b) | Balanced accuracy | 60.6% (57.4%–64.3%) | 9 | | Logistic regression on 100-ms means | Trained at 10 Hz (run a), tested at 5 Hz (run b) | Balanced accuracy | 67.5% (59.9%–75.2%) | 9 | | Logistic regression on 100-ms means | Trained at 10 Hz (run a), tested at 6 Hz (run b) | Balanced accuracy | 69.8% (61.9%–77.2%) | 9 | | Logistic regression on 100-ms means | Trained at 10 Hz (run a), tested at 10 Hz (run b) | Balanced accuracy | 63.0% (57.7%–68.0%) | 9 | | Logistic regression on 100-ms means | Trained at 5 Hz minus trained at 10 Hz, same 10-Hz test images | Paired difference | −2.2 pp (−6.8 pp–+2.8 pp) — 7 people lower, 2 higher. The interval crosses zero: no change is established. Rate and recording change together: not a causal effect of image rate. | 9 | ## Source and licence ### LTRSVP · EEG Signals from an RSVP Task **Credit** Ana Matran (University of Essex) · EEG Signals from an RSVP Task 1.0.0, PhysioNet (2017), doi:10.13026/C2KX0P. Original publication: Ana Matran-Fernandez and Riccardo Poli, Towards the automated localisation of targets in rapid image-sifting by collaborative brain-computer interfaces, PLoS ONE 12(5): e0178498 (2017), doi:10.1371/journal.pone.0178498. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) · [Dataset record ↗](https://physionet.org/content/ltrsvp/1.0.0/) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ltrsvp) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *LTRSVP · EEG Signals from an RSVP Task: EEG decoding results*. https://bci.report/datasets/ltrsvp/ Figures from release [`extension-update-20261002`](https://bci.report/releases/#extension-update-20261002) (2026-10-02). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ltrsvp/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Mobile BCI dataset (SSVEP and ERP paradigms): EEG decoding results SSVEP and ERP while standing, walking and running Scalp- and ear-EEG dataset from Korea University (Lee et al., Scientific Data, 2021): 24 participants (23 for SSVEP) performed ERP and SSVEP BCI tasks on a treadmill while standing, walking and slightly running (0 to 2.0 m/s), recorded with 32-channel scalp-EEG, 14-channel ear-EEG, EOG and IMUs. Original release on OSF; NEMAR nm000201 (ERP) and nm000125 (SSVEP) are BIDS derivatives. **Also known as** Mobile BCI dataset · Lee2021 · Lee2021Mobile · NEMAR nm000201 · NEMAR nm000125 · OSF r7s9b Description sources [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC8688416/) · [osf.io](https://osf.io/r7s9b/) · [nemar.org](https://nemar.org/dataset/nm000125) · [nemar.org](https://nemar.org/dataset/nm000201) · [doi.org](https://doi.org/10.82901/nemar.nm000125) · [doi.org](https://doi.org/10.82901/nemar.nm000201) ## Where it appears - [Does EEG decoding still work while walking or running?](https://bci.report/topics/on-the-move/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Movement · SSVEP, 5-second windows Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 12 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Author-style CCA · Standing · scalp · 8 ch: 88.7% (80.0%–95.9%) - Author-style CCA · Slow walk · 0.8 m/s · scalp · 8 ch: 82.9% (74.3%–90.4%) - Author-style CCA · Fast walk · 1.6 m/s · scalp · 8 ch: 80.9% (72.0%–88.8%) - Spectral ridge · Standing · scalp · 8 ch: 85.0% (77.0%–92.0%) - Spectral ridge · Slow walk · 0.8 m/s · scalp · 8 ch: 79.2% (70.1%–87.5%) - Spectral ridge · Fast walk · 1.6 m/s · scalp · 8 ch: 77.2% (67.7%–85.7%) - Author-style CCA · Standing · ear · 14 ch: 53.1% (47.4%–59.6%) - Author-style CCA · Slow walk · 0.8 m/s · ear · 14 ch: 43.4% (39.6%–47.8%) - Author-style CCA · Fast walk · 1.6 m/s · ear · 14 ch: 39.1% (36.7%–41.8%) - Spectral ridge · Standing · ear · 14 ch: 52.0% (46.9%–57.8%) - Spectral ridge · Slow walk · 0.8 m/s · ear · 14 ch: 45.4% (40.7%–51.2%) - Spectral ridge · Fast walk · 1.6 m/s · ear · 14 ch: 40.0% (36.4%–43.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Author-style CCA](https://bci.report/methods/cca/) | Standing · scalp · 8 ch | Balanced accuracy | 88.7% (80.0%–95.9%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 82.9% (74.3%–90.4%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 80.9% (72.0%–88.8%) | 23 | | Spectral ridge | Standing · scalp · 8 ch | Balanced accuracy | 85.0% (77.0%–92.0%) | 23 | | Spectral ridge | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 79.2% (70.1%–87.5%) | 23 | | Spectral ridge | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 77.2% (67.7%–85.7%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Standing · ear · 14 ch | Balanced accuracy | 53.1% (47.4%–59.6%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Slow walk · 0.8 m/s · ear · 14 ch | Balanced accuracy | 43.4% (39.6%–47.8%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Fast walk · 1.6 m/s · ear · 14 ch | Balanced accuracy | 39.1% (36.7%–41.8%) | 23 | | Spectral ridge | Standing · ear · 14 ch | Balanced accuracy | 52.0% (46.9%–57.8%) | 23 | | Spectral ridge | Slow walk · 0.8 m/s · ear · 14 ch | Balanced accuracy | 45.4% (40.7%–51.2%) | 23 | | Spectral ridge | Fast walk · 1.6 m/s · ear · 14 ch | Balanced accuracy | 40.0% (36.4%–43.7%) | 23 | ### Movement · SSVEP, 2-second windows Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 12 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Author-style CCA · Standing · scalp · 8 ch: 82.3% (74.6%–89.4%) - Author-style CCA · Slow walk · 0.8 m/s · scalp · 8 ch: 70.7% (62.3%–78.6%) - Author-style CCA · Fast walk · 1.6 m/s · scalp · 8 ch: 68.6% (60.1%–76.7%) - Spectral ridge · Standing · scalp · 8 ch: 78.7% (70.7%–86.1%) - Spectral ridge · Slow walk · 0.8 m/s · scalp · 8 ch: 69.9% (61.2%–78.1%) - Spectral ridge · Fast walk · 1.6 m/s · scalp · 8 ch: 68.6% (60.4%–76.3%) - LaBraM · Standing · scalp · 8 ch: 59.5% (53.6%–65.5%) - LaBraM · Slow walk · 0.8 m/s · scalp · 8 ch: 54.7% (49.5%–60.1%) - LaBraM · Fast walk · 1.6 m/s · scalp · 8 ch: 47.8% (43.3%–52.3%) - CBraMod · Standing · scalp · 8 ch: 59.9% (54.3%–65.4%) - CBraMod · Slow walk · 0.8 m/s · scalp · 8 ch: 46.2% (41.1%–51.4%) - CBraMod · Fast walk · 1.6 m/s · scalp · 8 ch: 46.0% (42.2%–49.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Author-style CCA](https://bci.report/methods/cca/) | Standing · scalp · 8 ch | Balanced accuracy | 82.3% (74.6%–89.4%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 70.7% (62.3%–78.6%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 68.6% (60.1%–76.7%) | 23 | | Spectral ridge | Standing · scalp · 8 ch | Balanced accuracy | 78.7% (70.7%–86.1%) | 23 | | Spectral ridge | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 69.9% (61.2%–78.1%) | 23 | | Spectral ridge | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 68.6% (60.4%–76.3%) | 23 | | [LaBraM](https://bci.report/methods/labram/) | Standing · scalp · 8 ch | Balanced accuracy | 59.5% (53.6%–65.5%) | 23 | | [LaBraM](https://bci.report/methods/labram/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 54.7% (49.5%–60.1%) | 23 | | [LaBraM](https://bci.report/methods/labram/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 47.8% (43.3%–52.3%) | 23 | | [CBraMod](https://bci.report/methods/cbramod/) | Standing · scalp · 8 ch | Balanced accuracy | 59.9% (54.3%–65.4%) | 23 | | [CBraMod](https://bci.report/methods/cbramod/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 46.2% (41.1%–51.4%) | 23 | | [CBraMod](https://bci.report/methods/cbramod/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 46.0% (42.2%–49.6%) | 23 | ### Movement · ERP Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Temporal-feature L2 logistic regression | Standing · scalp · 32 ch | ROC AUC | 0.859 (0.820–0.897) | 24 | | Temporal-feature L2 logistic regression | Slow walk · 0.8 m/s · scalp · 32 ch | ROC AUC | 0.740 (0.703–0.777) | 24 | | Temporal-feature L2 logistic regression | Fast walk · 1.6 m/s · scalp · 32 ch | ROC AUC | 0.635 (0.602–0.673) | 24 | | Temporal-feature L2 logistic regression | Running · 2.0 m/s · scalp · 32 ch | ROC AUC | 0.586 (0.561–0.613) | 17 | | Temporal-feature L2 logistic regression | Standing · ear · 14 ch | ROC AUC | 0.734 (0.694–0.774) | 24 | | Temporal-feature L2 logistic regression | Slow walk · 0.8 m/s · ear · 14 ch | ROC AUC | 0.628 (0.595–0.661) | 24 | | Temporal-feature L2 logistic regression | Fast walk · 1.6 m/s · ear · 14 ch | ROC AUC | 0.563 (0.540–0.587) | 24 | | Temporal-feature L2 logistic regression | Running · 2.0 m/s · ear · 14 ch | ROC AUC | 0.517 (0.491–0.541) | 17 | ## Source and licence ### Mobile BCI dataset (SSVEP and ERP paradigms) **Credit** Lee, Y., Shin, G., Lee, M., & Lee, S. (2026). Lee2021 – SSVEP paradigm of the Mobile BCI dataset (Version v1.0.2) [Data set]. NEMAR. https://doi.org/10.82901/nemar.nm000125 · Lee, Y., Shin, G., Lee, M., & Lee, S. (2026). ERP paradigm of the Mobile BCI dataset (Version v1.0.2) [Data set]. NEMAR. https://doi.org/10.82901/nemar.nm000201 CC BY 4.0 · [Dataset record ↗](https://nemar.org/dataset/nm000125) · [nemar.org ↗](https://nemar.org/dataset/nm000201) · [doi.org ↗](https://doi.org/10.1038/s41597-021-01094-4) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-mobile-bci) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Mobile BCI dataset (SSVEP and ERP paradigms): EEG decoding results*. https://bci.report/datasets/mobile-bci/ Figures from release [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/mobile-bci/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # MC_RTT (Neural Latents Benchmark): intracortical spiking results One macaque’s motor-cortex spiking during self-paced reaching: finger velocity MC_RTT (DANDI 000129) holds sorted spiking activity and behaviour from one rhesus macaque making self-paced reaches between targets drawn at random from an 8 × 8 grid, with no gaps or pre-movement delays, recorded by an electrode array implanted in primary motor cortex, together with finger, cursor and target position. Recorded by Joseph O'Doherty and colleagues in the Sabes Lab at UCSF, prepared for the Neural Latents Benchmark (Pei et al., 2021) and released on DANDI under CC BY 4.0. Only its stored training and validation trials are used here; its test file is not opened. **Also known as** MC_RTT · Neural Latents Benchmark MC_RTT · DANDI 000129 · macaque motor cortex spiking activity during self-paced reaching · O'Doherty et al. random target task Description sources [doi.org](https://doi.org/10.48324/dandi.000129/0.241017.1444) · [arxiv.org](https://arxiv.org/abs/2109.04463) · [doi.org](https://doi.org/10.5281/zenodo.3854034) ## Where it appears - [Do simple baselines work on signals other than scalp EEG?](https://bci.report/topics/beyond-scalp-eeg/#mc-rtt) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Beyond scalp EEG · one macaque’s motor-cortex spiking, finger velocity within one session Source page: [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/#mc-rtt) · [beyond-scalp-update.json](https://bci.report/data/beyond-scalp-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Ridge on spike-count history | Held-out trials of the same session · x/y finger velocity | R², mean of x and y (higher is better) | 0.240 | 1 macaque (animal data) | | Training mean — comparator | Held-out trials of the same session · x/y finger velocity | R², mean of x and y (higher is better) | −0.00175 | 1 macaque (animal data) | | Ridge on spike-count history | Held-out trials of the same session · x/y finger velocity | R², x (higher is better) | 0.202 | 1 macaque (animal data) | | Training mean — comparator | Held-out trials of the same session · x/y finger velocity | R², x (higher is better) | −0.000104 | 1 macaque (animal data) | | Ridge on spike-count history | Held-out trials of the same session · x/y finger velocity | R², y (higher is better) | 0.278 | 1 macaque (animal data) | | Training mean — comparator | Held-out trials of the same session · x/y finger velocity | R², y (higher is better) | −0.00340 | 1 macaque (animal data) | | Ridge on spike-count history | Held-out trials of the same session · x/y finger velocity | RMSE in m/s (lower is better) | 0.056 | 1 macaque (animal data) | | Training mean — comparator | Held-out trials of the same session · x/y finger velocity | RMSE in m/s (lower is better) | 0.064 | 1 macaque (animal data) | | Ridge on spike-count history | Held-out trials of the same session · x/y finger velocity | MAE in m/s (lower is better) | 0.038 — The ridge is worse than the training mean on this metric. | 1 macaque (animal data) | | Training mean — comparator | Held-out trials of the same session · x/y finger velocity | MAE in m/s (lower is better) | 0.033 | 1 macaque (animal data) | One macaque in one session, with training and validation trials interleaved: with one animal every figure is that animal’s, and no interval is reported, because the validation rows are correlated samples, not independent subjects. Offline finger-velocity decoding, scored by this site; the Neural Latents Benchmark’s co-smoothing metric was not computed. ## Source and licence ### MC_RTT (Neural Latents Benchmark) **Credit** O'Doherty, Joseph (2024) MC_RTT: macaque motor cortex spiking activity during self-paced reaching (Version 0.241017.1444) [Data set]. DANDI archive. https://doi.org/10.48324/dandi.000129/0.241017.1444, CC BY 4.0. Recorded by Joseph E. O'Doherty, Mariana M. B. Cardoso, Joseph G. Makin and Philip N. Sabes in the Sabes Lab at UCSF (Nonhuman Primate Reaching with Multichannel Sensorimotor Cortex Electrophysiology, Zenodo, doi:10.5281/zenodo.3854034; procedure: Makin et al., Journal of Neural Engineering 15(2), 026010 (2018), doi:10.1088/1741-2552/aa9e95). Prepared for the Neural Latents Benchmark: Felix Pei, Joel Ye et al., Neural Latents Benchmark '21: Evaluating latent variable models of neural population activity, NeurIPS 2021 Track on Datasets and Benchmarks, arXiv:2109.04463. Derived analysis by BCI Report; not endorsed by the authors. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.48324/dandi.000129/0.241017.1444) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-nlb-mc-rtt) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *MC_RTT (Neural Latents Benchmark): intracortical spiking results*. https://bci.report/datasets/nlb-mc-rtt/ Figures from release [`beyond-scalp-update-20261012`](https://bci.report/releases/#beyond-scalp-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/nlb-mc-rtt/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # OpenBMI motor imagery (Lee et al. 2019): EEG decoding results Motor imagery: left or right hand across two sessions, and imagery or rest The OpenBMI dataset (GigaScience Database 100542) holds EEG from 54 healthy participants, each recorded in two sessions on different days in three BCI paradigms — motor imagery, event-related potentials and SSVEP — with 62 electrodes at 1,000 Hz, plus questionnaires, resting-state and artifact recordings. Released by Lee et al. (Korea University) with the OpenBMI toolbox and described in GigaScience (2019). Only the motor-imagery training runs are used here. **Also known as** OpenBMI · Lee2019 · Lee2019_MI · GigaDB 100542 · EEG Dataset and OpenBMI Toolbox for Three BCI Paradigms · Lee et al. (2019) Description sources [doi.org](https://doi.org/10.5524/100542) · [doi.org](https://doi.org/10.1093/gigascience/giz002) · [api.datacite.org](https://api.datacite.org/dois/10.5524/100542) ## Where it appears - [How much calibration data does an EEG decoder need?](https://bci.report/topics/calibration-budget/) - [Can one EEG model answer several questions about the same data as well as separate models, and at what cost?](https://bci.report/topics/shared-encoder/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Calibration budget · same person, next session, motor imagery Source page: [How much calibration?](https://bci.report/topics/calibration-budget/) · Chance level 50.0% · [large-source-update.json](https://bci.report/data/large-source-update.json) **Balanced accuracy, 8 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Log-covariance + shrinkage LDA · Session 1 → session 2 · 0 labeled session-2 trials: 67.9% (63.6%–72.4%) - Log-covariance + shrinkage LDA · Session 1 → session 2 · 10 labeled session-2 trials: 69.1% (64.8%–73.5%) - Log-covariance + shrinkage LDA · Session 1 → session 2 · 20 labeled session-2 trials: 69.2% (65.3%–73.4%) - Log-covariance + shrinkage LDA · Session 1 → session 2 · 40 labeled session-2 trials: 72.4% (68.9%–76.2%) - Relative PSD + standardized ridge · Session 1 → session 2 · 0 labeled session-2 trials: 57.3% (54.4%–60.5%) - Relative PSD + standardized ridge · Session 1 → session 2 · 10 labeled session-2 trials: 57.4% (54.5%–60.5%) - Relative PSD + standardized ridge · Session 1 → session 2 · 20 labeled session-2 trials: 57.9% (55.2%–60.8%) - Relative PSD + standardized ridge · Session 1 → session 2 · 40 labeled session-2 trials: 59.0% (56.0%–62.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Log-covariance + shrinkage LDA | Session 1 → session 2 · 0 labeled session-2 trials | Balanced accuracy | 67.9% (63.6%–72.4%) | 51 | | Log-covariance + shrinkage LDA | Session 1 → session 2 · 10 labeled session-2 trials | Balanced accuracy | 69.1% (64.8%–73.5%) | 51 | | Log-covariance + shrinkage LDA | Session 1 → session 2 · 20 labeled session-2 trials | Balanced accuracy | 69.2% (65.3%–73.4%) | 51 | | Log-covariance + shrinkage LDA | Session 1 → session 2 · 40 labeled session-2 trials | Balanced accuracy | 72.4% (68.9%–76.2%) | 51 | | Log-covariance + shrinkage LDA | 10 minus 0 labeled session-2 trials, the same people and test trials | Paired difference | +1.2 pp (−0.6 pp–+3.3 pp) — 21 of 51 people declined, 8 by 5 points or more. The interval includes zero: the gain is not established. | 51 | | Log-covariance + shrinkage LDA | 20 minus 0 labeled session-2 trials, the same people and test trials | Paired difference | +1.4 pp (−0.3 pp–+3.2 pp) — 24 of 51 people declined, 6 by 5 points or more. The interval includes zero: the gain is not established. | 51 | | Log-covariance + shrinkage LDA | 40 minus 0 labeled session-2 trials, the same people and test trials | Paired difference | +4.6 pp (+2.3 pp–+6.9 pp) — 13 of 51 people declined, 6 by 5 points or more. | 51 | | Relative PSD + standardized ridge | Session 1 → session 2 · 0 labeled session-2 trials | Balanced accuracy | 57.3% (54.4%–60.5%) | 51 | | Relative PSD + standardized ridge | Session 1 → session 2 · 10 labeled session-2 trials | Balanced accuracy | 57.4% (54.5%–60.5%) | 51 | | Relative PSD + standardized ridge | Session 1 → session 2 · 20 labeled session-2 trials | Balanced accuracy | 57.9% (55.2%–60.8%) | 51 | | Relative PSD + standardized ridge | Session 1 → session 2 · 40 labeled session-2 trials | Balanced accuracy | 59.0% (56.0%–62.1%) | 51 | | Relative PSD + standardized ridge | 10 minus 0 labeled session-2 trials, the same people and test trials | Paired difference | +0.1 pp (−1.1 pp–+1.2 pp) — 25 of 51 people declined, 6 by 5 points or more. The interval includes zero: the gain is not established. | 51 | | Relative PSD + standardized ridge | 20 minus 0 labeled session-2 trials, the same people and test trials | Paired difference | +0.6 pp (−0.9 pp–+2.0 pp) — 19 of 51 people declined, 10 by 5 points or more. The interval includes zero: the gain is not established. | 51 | | Relative PSD + standardized ridge | 40 minus 0 labeled session-2 trials, the same people and test trials | Paired difference | +1.7 pp (−0.1 pp–+3.5 pp) — 16 of 51 people declined, 6 by 5 points or more. The interval includes zero: the gain is not established. | 51 | ### One model, several questions · motor imagery (route 2) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · Chance level 50.0% · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) **Balanced accuracy, 14 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - EEGNet · fixed heads (B-lin) · imagery or rest (MI-A) · trained from scratch, three seeds: 73.2% (70.2%–76.2%) - EEGNet · shared hidden layer (B-sh) · imagery or rest (MI-A) · trained from scratch, three seeds: 74.3% (71.3%–77.3%) - EEGNet · question-conditioned head (C1) · imagery or rest (MI-A) · trained from scratch, three seeds: 74.6% (71.9%–77.5%) - EEGNet · separate model (A) · imagery or rest (MI-A) · trained from scratch, three seeds: 74.2% (71.3%–76.9%) - CBraMod · fixed heads (B-lin-fz-sgd) · imagery or rest (MI-A) · heads on frozen features, three seeds: 69.2% (67.0%–71.5%) - CBraMod · shared hidden layer (B-sh-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds: 70.9% (68.4%–73.4%) - CBraMod · question-conditioned head (C1-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds: 70.8% (68.4%–73.2%) - EEGNet · fixed heads (B-lin) · which hand (MI-B) · trained from scratch, three seeds: 67.5% (64.6%–70.5%) - EEGNet · shared hidden layer (B-sh) · which hand (MI-B) · trained from scratch, three seeds: 67.0% (64.1%–69.9%) - EEGNet · question-conditioned head (C1) · which hand (MI-B) · trained from scratch, three seeds: 67.4% (64.4%–70.5%) - EEGNet · separate model (A) · which hand (MI-B) · trained from scratch, three seeds: 69.0% (65.9%–72.0%) - CBraMod · fixed heads (B-lin-fz-sgd) · which hand (MI-B) · heads on frozen features, three seeds: 56.1% (54.8%–57.4%) - CBraMod · shared hidden layer (B-sh-fz) · which hand (MI-B) · heads on frozen features, three seeds: 56.7% (55.4%–58.2%) - CBraMod · question-conditioned head (C1-fz) · which hand (MI-B) · heads on frozen features, three seeds: 57.2% (55.8%–58.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 73.2% (70.2%–76.2%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 74.3% (71.3%–77.3%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 74.6% (71.9%–77.5%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 74.2% (71.3%–76.9%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | imagery or rest (MI-A) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +1.48 pp (+0.09–+2.89 pp) — C1 higher; 2 pp margin not met; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | imagery or rest (MI-A) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | +0.31 pp (−0.91–+1.54 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | imagery or rest (MI-A) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(MI-A)) | Paired difference, balanced accuracy | −1.00 pp (−2.21–+0.19 pp) — No difference shown; 2 pp margin not met: inconclusive at this sample size; gate passed. | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · imagery or rest (MI-A) · heads on frozen features, three seeds | Balanced accuracy | 69.2% (67.0%–71.5%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds | Balanced accuracy | 70.9% (68.4%–73.4%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds | Balanced accuracy | 70.8% (68.4%–73.2%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | imagery or rest (MI-A) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +1.54 pp (+0.86–+2.30 pp) — C1-fz higher; 2 pp margin not met; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 67.5% (64.6%–70.5%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 67.0% (64.1%–69.9%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 67.4% (64.4%–70.5%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 69.0% (65.9%–72.0%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | which hand (MI-B) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | −0.09 pp (−1.24–+1.04 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | which hand (MI-B) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | +0.45 pp (−0.60–+1.34 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | which hand (MI-B) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(MI-B)) | Paired difference, balanced accuracy | −1.43 pp (−2.57–−0.39 pp) — A(MI-B) higher; 2 pp margin not met; gate passed. | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · which hand (MI-B) · heads on frozen features, three seeds | Balanced accuracy | 56.1% (54.8%–57.4%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · which hand (MI-B) · heads on frozen features, three seeds | Balanced accuracy | 56.7% (55.4%–58.2%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · which hand (MI-B) · heads on frozen features, three seeds | Balanced accuracy | 57.2% (55.8%–58.6%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | which hand (MI-B) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +1.09 pp (+0.08–+2.12 pp) — C1-fz higher; 2 pp margin not met (not read: at floor); at floor: reported, not counted. | 51 | **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. The OpenBMI numbers here use route 2's own 2-s windows, label-free high-pass and person-disjoint folds; they are not comparable with the site's OpenBMI cross-session calibration figures. A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ## Source and licence ### OpenBMI motor imagery (Lee et al. 2019) **Credit** Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.5524/100542) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-openbmi) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *OpenBMI motor imagery (Lee et al. 2019): EEG decoding results*. https://bci.report/datasets/openbmi/ Figures from releases [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007) (2026-10-07), [`large-source-update-20261003`](https://bci.report/releases/#large-source-update-20261003) (2026-10-03). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/openbmi/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Orthosis error-detection dataset (Kueper et al. 2024): EEG decoding results Injected orthosis device errors against standard device events, EEG only: one cohort in three configurations The EEG and EMG dataset for the detection of errors introduced by an active orthosis device (Zenodo 8345429) holds recordings from eight healthy subjects whose right arm an active orthosis moved in elbow flexion and extension; during some movements an error was deliberately introduced, and the orthosis briefly moved the opposite way. EEG was recorded with a 64-channel system at 500 Hz, alongside eight-channel EMG. Prepared for the IJCAI 2023 competition CC6 (IntEr-HRI) and released in three versions, the third the complete dataset, by Niklas Kueper and colleagues (DFKI and the University of Duisburg-Essen) under CC BY 4.0; described in Frontiers in Human Neuroscience (2024). Only the EEG is used here. **Also known as** EEG and EMG dataset for the detection of errors introduced by an active orthosis device · IJCAI'23 CC6 IntEr-HRI competition dataset · Zenodo 8345429 · Kueper et al. (2024) Description sources [doi.org](https://doi.org/10.5281/zenodo.8345429) · [doi.org](https://doi.org/10.3389/fnhum.2024.1304311) · [zenodo.org](https://zenodo.org/api/records/8345429) ## Where it appears - [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/#orthosis) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Pilot studies · orthosis error events, the 64-channel classical baseline Source page: [Pilot studies](https://bci.report/topics/pilot-studies/#orthosis) · Chance level 50.0% · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Fixed EEG features + logistic regression · 64 channels: 73.0% (68.0%–77.7%) - Direction/time confound probe (no EEG): 51.4% (50.4%–52.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Fixed EEG features + logistic regression · 64 channels | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 73.0% (68.0%–77.7%) | 8 | | Fixed EEG features + logistic regression · 64 channels | A person left out of training · 0–0.5 s after each recorded device event | AUROC | 0.812 (0.751–0.863) | 8 | | Fixed EEG features + logistic regression · 64 channels | A person left out of training · 0–0.5 s after each recorded device event | Average precision | 0.585 (0.492–0.672) | 8 | | Direction/time confound probe (no EEG) — comparator | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 51.4% (50.4%–52.2%) | 8 | | Training class prior — comparator | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 50.0% | 8 | | Fixed EEG features + logistic regression · 64 channels | Minus the confound probe, matched person by person | Paired difference | +21.6 pp — No interval was computed for this difference. EEG was above the confound probe for 8 of 8 people. | 8 | | Permutation control: training labels shuffled within person and run · Fixed EEG features + logistic regression · 64 channels — comparator | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 50.1% — The mean of 20 replicates, whose means ranged from 48.0% to 52.6%: an observed range, not a confidence interval. | 8 | One cohort — 8 people, 83 runs, 2,407 recorded device events, 474 of them injected errors — published as three configurations and counted once. The labels are recorded device events: the results do not show that a person recognised an error, and movement, muscle or device artifacts may contribute. Errors are a minority of events, so average precision is read against that share. A response-clean subset is counts only: 451 events, 446 errors and 5 standard events, with 3 people having both classes, so its scores are not published. ### Pilot studies · orthosis error events, the matched 58-channel classical configuration (the same cohort) Source page: [Pilot studies](https://bci.report/topics/pilot-studies/#orthosis) · Chance level 50.0% · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Fixed EEG features + logistic regression · 58 channels: 71.0% (64.8%–76.9%) - Direction/time confound probe (no EEG): 51.4% (50.4%–52.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Fixed EEG features + logistic regression · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 71.0% (64.8%–76.9%) | 8 | | Fixed EEG features + logistic regression · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | AUROC | 0.793 (0.719–0.854) | 8 | | Fixed EEG features + logistic regression · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | Average precision | 0.544 (0.438–0.654) | 8 | | Direction/time confound probe (no EEG) — comparator | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 51.4% (50.4%–52.3%) | 8 | | Permutation control: training labels shuffled within person and run · Fixed EEG features + logistic regression · 58 channels — comparator | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 50.1% — The mean of 20 replicates, whose means ranged from 48.2% to 51.7%: an observed range, not a confidence interval. | 8 | The same cohort, labels and caveats as the 64-channel baseline above, counted once. The 58 channels are the ones EEGPT supports, fixed before scoring. A second configuration of the same people, not a replication: the 64-channel scores were seen before it was frozen, and no conclusion about how many channels are needed follows. ### Pilot studies · orthosis error events, frozen EEGPT against the matched classical features (58 channels, the same cohort) Source page: [Pilot studies](https://bci.report/topics/pilot-studies/#orthosis) · Chance level 50.0% · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - EEGPT · frozen encoder + the same supervised probe · 58 channels: 70.3% (66.2%–73.9%) - Fixed EEG features + logistic regression · 58 channels: 71.0% (64.8%–76.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | EEGPT · frozen encoder + the same supervised probe · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 70.3% (66.2%–73.9%) | 8 | | EEGPT · frozen encoder + the same supervised probe · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | AUROC | 0.780 (0.721–0.827) | 8 | | EEGPT · frozen encoder + the same supervised probe · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | Average precision | 0.564 (0.489–0.630) | 8 | | Fixed EEG features + logistic regression · 58 channels | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 71.0% (64.8%–76.9%) | 8 | | EEGPT · frozen encoder + the same supervised probe · 58 channels | Minus the matched classical features, the same people and folds | Paired difference | −0.6 pp (−4.0–+3.2 pp) — The paired interval spans zero: neither arm is shown to be better, and the two are not shown to be equivalent. | 8 | | Permutation control: training labels shuffled within person and run · EEGPT · frozen encoder + the same supervised probe · 58 channels — comparator | A person left out of training · 0–0.5 s after each recorded device event | Balanced accuracy | 49.8% — The mean of 20 replicates, whose means ranged from 47.6% to 52.2%: an observed range, not a confidence interval. | 8 | The same cohort, channels, folds and probe as the 58-channel configuration above, counted once. Supervised transfer: the frozen encoder’s outputs feed a probe trained on the other people’s labels, not zero-shot. Overlap between this dataset and EEGPT’s pretraining data has not been excluded, and the 64-channel classical scores were seen before this comparison was frozen. EEGPT’s checkpoint licence is unresolved and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. Not evidence about EEG foundation models in general. ## Source and licence ### Orthosis error-detection dataset (Kueper et al. 2024) **Credit** Niklas Kueper, Kartik Chari, Judith Bütefür, Julia Habenicht, Tobias Rossol, Su Kyoung Kim, Marc Tabie, Frank Kirchner and Elsa Andrea Kirchner · EEG and EMG dataset for the detection of errors introduced by an active orthosis device (IJCAI’23 CC6 Competition), Zenodo, doi:10.5281/zenodo.8345429, CC BY 4.0. Paper: Niklas Kueper et al., EEG and EMG dataset for the detection of errors introduced by an active orthosis device, Frontiers in Human Neuroscience 18 (2024), doi:10.3389/fnhum.2024.1304311. Derived analysis by BCI Report; not endorsed by the authors. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.5281/zenodo.8345429) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-orthosis-error-detection) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Orthosis error-detection dataset (Kueper et al. 2024): EEG decoding results*. https://bci.report/datasets/orthosis-error-detection/ Figures from release [`pilot-studies-update-20261012`](https://bci.report/releases/#pilot-studies-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/orthosis-error-detection/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Phantom EEG with motion, muscle and eye artifacts: EEG decoding results Physical head phantom with motion, muscle and eye artifacts 128-channel EEG recorded by Downey and Ferris (University of Florida) from an electrically conductive head phantom, with no human participants, containing 10 brain sources and 10 artifact sources, under six conditions: brain only; brain plus eye, neck-muscle, facial-muscle or walking-motion artifacts; or all together. Released on OpenNeuro as ds004784 (CC0, 2023) alongside the iCanClean paper (Sensors, 2023). **Also known as** Phantom EEG Dataset with Motion, Muscle, and Eye Artifacts and Example Scripts · OpenNeuro ds004784 · Phantom EEG with motion, muscle and eye artifacts Description sources [openneuro.org](https://openneuro.org/datasets/ds004784) · [doi.org](https://doi.org/10.18112/openneuro.ds004784.v1.0.4) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC10574843/) ## Where it appears - [Does EEG decoding still work while walking or running?](https://bci.report/topics/on-the-move/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Movement · physical head phantom (engineering check) Source page: [On the move](https://bci.report/topics/on-the-move/) · [evidence-update.json](https://bci.report/data/evidence-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Fixed multi-output ridge regression | Brain | Signed correlation r | 0.761 | no human participants | | Fixed multi-output ridge regression | Brain | Predictive R² | 0.567 | no human participants | | Fixed multi-output ridge regression | Eyes | Signed correlation r | 0.492 | no human participants | | Fixed multi-output ridge regression | Eyes | Predictive R² | −33.612 | no human participants | | Fixed multi-output ridge regression | Walking | Signed correlation r | 0.549 | no human participants | | Fixed multi-output ridge regression | Walking | Predictive R² | −795.772 | no human participants | | Fixed multi-output ridge regression | Neck | Signed correlation r | 0.310 | no human participants | | Fixed multi-output ridge regression | Neck | Predictive R² | −596.820 | no human participants | | Fixed multi-output ridge regression | Facial | Signed correlation r | 0.143 | no human participants | | Fixed multi-output ridge regression | Facial | Predictive R² | −164.313 | no human participants | | Fixed multi-output ridge regression | All | Signed correlation r | 0.115 | no human participants | | Fixed multi-output ridge regression | All | Predictive R² | −142.832 | no human participants | ## Source and licence ### Phantom EEG with motion, muscle and eye artifacts **Credit** Ryan J. Downey and Daniel P. Ferris · Phantom EEG Dataset with Motion, Muscle, and Eye Artifacts and Example Scripts, OpenNeuro ds004784 v1.0.4, doi:10.18112/openneuro.ds004784.v1.0.4. Method: doi:10.3390/s23198214. [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds004784.v1.0.4) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-phantom) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Phantom EEG with motion, muscle and eye artifacts: EEG decoding results*. https://bci.report/datasets/phantom/ Figures from release [`evidence-update-20260922`](https://bci.report/releases/#evidence-update-20260922) (2026-09-22). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/phantom/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # REFED emotion dataset (EEG-fNIRS): EEG decoding results Self-reported valence and arousal while watching videos: a stored-clip diagnostic REFED (A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset) holds simultaneously recorded EEG and fNIRS from 32 healthy participants watching 15 emotion-eliciting video clips, with valence and arousal annotated continuously, in real time, by the participants themselves. Released on Hugging Face by Xiaojun Ning, Ziyu Jia and colleagues under CC BY-NC-SA 4.0, behind a gate that asks users to agree to non-commercial use and to cause no harm to human subjects, and described at NeurIPS 2025 (Datasets and Benchmarks Track). Only stored EEG clips from two nominal participant folders are used here. **Also known as** REFED · A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset · REFED2025/REFED-dataset · Ning et al. (2025) Description sources [huggingface.co](https://huggingface.co/datasets/REFED2025/REFED-dataset) · [papers.neurips.cc](https://papers.neurips.cc/paper_files/paper/2025/file/2bf0ed7c35d9d84128e7f7c72ab76402-Paper-Datasets_and_Benchmarks_Track.pdf) ## Where it appears - [What can a small pilot show, and what can it not?](https://bci.report/topics/pilot-studies/#refed) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Pilot studies · REFED stored clips, a computational diagnostic Source page: [Pilot studies](https://bci.report/topics/pilot-studies/#refed) · [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | EEG Ridge (relative band power) | Shared videos, a folder held out (familiar stimuli) | Equal-axis mean absolute error, raw self-report scale (lower is better) | 107.182 | 2 nominal participant folders | | Training-set mean — comparator | Shared videos, a folder held out (familiar stimuli) | Equal-axis mean absolute error, raw self-report scale (lower is better) | 40.287 | 2 nominal participant folders | | Same-video training lookup — comparator | Shared videos, a folder held out (familiar stimuli) | Equal-axis mean absolute error, raw self-report scale (lower is better) | 27.325 | 2 nominal participant folders | | Duration / numeric-video-ID ridge — comparator | Shared videos, a folder held out (familiar stimuli) | Equal-axis mean absolute error, raw self-report scale (lower is better) | 37.910 | 2 nominal participant folders | | EEG Ridge (relative band power) | Folder and video IDs held out together | Equal-axis mean absolute error, raw self-report scale (lower is better) | 91.984 | 2 nominal participant folders | | Training-set mean — comparator | Folder and video IDs held out together | Equal-axis mean absolute error, raw self-report scale (lower is better) | 41.032 | 2 nominal participant folders | | Same-video training lookup — comparator | Folder and video IDs held out together | Equal-axis mean absolute error, raw self-report scale (lower is better) | 41.032 | 2 nominal participant folders | | Duration / numeric-video-ID ridge — comparator | Folder and video IDs held out together | Equal-axis mean absolute error, raw self-report scale (lower is better) | 47.098 | 2 nominal participant folders | 2 nominal participant folders with the same 15 videos: 30 clips, not as many independent people, so no interval or significance test is reported. EEG Ridge has a larger error than every source-only control in both conditions. With shared videos, the same-video lookup has the smallest error, because familiar stimuli carry the prediction; held out together, it has no matching video and equals the training-set mean. Targets are the mean of each clip’s stored valence and arousal labels on the published raw scale, not SAM scores. Stored EEG units, the EEG-to-rating clock and the crop end points are unresolved, and the rating means were seen before this design. ## Source and licence ### REFED emotion dataset (EEG-fNIRS) **Credit** Xiaojun Ning, Jing Wang, Zhiyang Feng, Tianzuo Xin, Shuo Zhang, Shaoqi Zhang, Zheng Lian, Yi Ding, Youfang Lin and Ziyu Jia · REFED: A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset, NeurIPS 2025 Datasets and Benchmarks Track. Data: REFED2025/REFED-dataset on Hugging Face, revision 0298355c2823946ac5800b2ec12a082bf2dcd01b, CC BY-NC-SA 4.0, gated. Derived analysis by BCI Report; not endorsed by the authors. [CC BY-NC-SA 4.0 ↗](https://creativecommons.org/licenses/by-nc-sa/4.0/) · [Dataset record ↗](https://huggingface.co/datasets/REFED2025/REFED-dataset) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-refed) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *REFED emotion dataset (EEG-fNIRS): EEG decoding results*. https://bci.report/datasets/refed/ Figures from release [`pilot-studies-update-20261012`](https://bci.report/releases/#pilot-studies-update-20261012) (2026-10-12). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/refed/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Every EEG dataset we know of, and how far we checked it Open register Every EEG dataset we know of, with the furthest step we have checked and what that step showed. Anything can be listed here. The score tables keep their own rules: rights review, aggregate results only, no redistribution. ## What the register holds Every number here is a value of [the register file](https://bci.report/data/open-register-20261106.json). - **Rows.** 488 rows on the dataset ladder: 293 for uploads on Hugging Face, 1 for OpenNeuro via EEGDash, 52 for this site’s datasets, 136 for MOABB records and 6 for corpora that require an agreement or request. By the furthest step checked: 435 at step 1, 4 at step 2, 10 at step 3, 32 at step 4 and 7 at step 5. The model ladder has another 62 rows, on [the model register](https://bci.report/methods/register/). The rows here were checked on 19 October 2026, 20 October 2026, 28 October 2026 and 6 November 2026. - **Hugging Face.** Read on 19 October 2026 at 14:23 UTC for 28 search terms and 20 tag filters: 1,922 listings. 737 are EEGDash’s mirrors of OpenNeuro and NEMAR records, one row here. Of the other 1,185, a keyword filter set aside 536 and the register’s classifier read 649, marking 400 as brain-signal data. A re-check then took out 5 under the EMG rule, 7 with simulated signals and 20 whose name was their only sign of brain signals. Of the 368 left, 278 are rows of their own, 79 sit in 15 rows for an uploader’s series, 5 are merged into this site’s rows, 1 is a MOABB record’s own upload and sits in that record’s row, and 5 are held for review. 146 of them carry no licence tag; 17 are gated with each request approved by the uploader, and 7 with automatic approval. - **EEGDash.** 737 mirrors: 546 of OpenNeuro records, 168 of NEMAR records and 23 of other records. 13 of this site’s rows and 6 MOABB rows have one. - **MOABB.** Release 1.7.2 has 206 loaders over 158 source records; 22 of those records are datasets this site already lists, merged into its rows. Catalog read on 19 October 2026. - **Corpora under agreement.** 6 corpora require a signed agreement, a request or a registration; the terms of all 6 are quoted from captured copies of their pages. The Hugging Face search terms and tags Search terms: `eeg`, `bci`, `brain-computer`, `brain computer interface`, `ecog`, `ieeg`, `seeg`, `intracranial`, `fnirs`, `nirs`, `ssvep`, `p300`, `motor imagery`, `motor-imagery`, `erp eeg`, `polysomnography`, `sleep-edf`, `sleep staging`, `neural decoding`, `spike sorting`, `neuropixels`, `emg`, `electroencephalography`, `brainwave`, `neuro signal`, `openbci`, `muse eeg`, `emotiv`. Tags: `eeg`, `EEG`, `bci`, `brain-computer-interface`, `neuroscience`, `electroencephalography`, `biosignals`, `biosignal`, `emg`, `ecog`, `ieeg`, `fnirs`, `meg`, `sleep`, `neural-decoding`, `brain`, `neuro`, `ssvep`, `motor-imagery`, `time-series-forecasting eeg`. ## The steps A dataset climbs five steps. A row’s step is the furthest one whose own evidence is on file; an item missing from an earlier step is listed beside the badge as a gap. The evidence for each step, item by item with its source, is in the JSON file. - Step 1 · Listed: Host record read: host, licence as stated and by whom, size, whether access is gated - Step 2 · Read: Files opened; a specification sheet from the files, compared with the card - Step 3 · Traced: Derivation, original terms against declared terms, consent or ethics statement, pinned revision - Step 4 · Run: Fixed-protocol baseline here, chance from trial counts, share of participants above chance, result or hold on a page - Step 5 · In the matrix: Comparable on a protocol page ### Rules - Gated data, or data that needs a signed or requested agreement, stays at step 1 unless we accept its terms. - Suspected re-uploads stay at step 1 whatever their gating. They are compared by metadata only, and none of their files is downloaded. - A finding about a re-upload is stated only when an independent check confirms it and a captured page of the original terms backs it. - Rows state facts. - Every row has a correction route. To correct a row, use the mail link at its end: the subject names the row’s id. How corrections are handled is on [the data-use page](https://bci.report/data-use/#register-corrections). ## Every row Each row links its host and gives its terms as stated and by whom, the step it reached with any gaps, and what that step showed. | Dataset | Signal · task | Terms as stated | Checked to | What we found | | --- | --- | --- | --- | --- | | **Hugging Face uploads** · 293 rows — Uploads the register’s classifier marked as brain-signal data in the snapshot counted above, each read from its listing and from its metadata at a pinned revision. A series of repositories from one uploader is one row. What the cards state about signal, task and participants is still being reviewed, so signal and task are left blank here. | | | | | | [2Levi/EEG_fNIRS_4_BCI ↗](https://huggingface.co/datasets?search=EEG_fNIRS_4_BCI) — Hugging Face · 12 files · 70,974,219,330 bytes · revision `0b2a3ec` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches Buccino, Keles and Omurtag 2016 hybrid EEG-fNIRS SMR BCI data in archive names and byte counts. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://doi.org/10.6084/m9.figshare.1619640.v1)).; Matches FineMI (EEG and fNIRS, eight upper-limb motor imagery tasks) in archive names and byte counts. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://doi.org/10.6084/m9.figshare.24123303)).; Matches HEFMI-ICH (hybrid EEG-fNIRS motor imagery, intracerebral haemorrhage) in archive names. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://doi.org/10.6084/m9.figshare.28955456.v4)).; Matches MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset (PhysioNet mimic-iv-ecg 1.0) in archive names and byte counts. That source’s terms page gives `odbl-1.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/mimic-iv-ecg/1.0/)).; Matches TDBRAIN (Two Decades Brainclinics Research Archive for Insights in Neurophysiology) in archive names. That source requires registration; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://brainclinics.com/data-use-agreement)).; Matches TU Berlin open-access simultaneous EEG and NIRS BCI dataset (Shin et al. 2017) in archive names. That source’s terms page gives `gpl`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](http://doc.ml.tu-berlin.de/hBCI/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0001) | | [Aditharavind/laya-neuro-decisions ↗](https://huggingface.co/datasets?search=laya-neuro-decisions) — Hugging Face · 9 files · 38,408,782 bytes · revision `816d527` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0002) | | [Alidr79/cueless_EEG_subject_identification ↗](https://huggingface.co/datasets?search=cueless_EEG_subject_identification) — Hugging Face · 271 files · 6,729,108,327 bytes · revision `5f4ba48` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0003) | | [Amirrzh/MI_EEG_Datasets ↗](https://huggingface.co/datasets?search=MI_EEG_Datasets) — Hugging Face · 4,817 files · 106,747,619,329 bytes · revision `738bfe6` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0004) | | [AnantaJ/BETA-from_Tsinghua_BCI_Lab ↗](https://huggingface.co/datasets?search=BETA-from_Tsinghua_BCI_Lab) — Hugging Face · 10 files · 5,278,929,409 bytes · revision `1762f4c` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0005) | | [AnantaJ/Benchmark-from_Tsinghua_BCI_Lab ↗](https://huggingface.co/datasets?search=Benchmark-from_Tsinghua_BCI_Lab) — Hugging Face · 38 files · 3,695,413,419 bytes · revision `1f8db8b` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0006) | | [AnasElkhabbaz/EMG-Dataset ↗](https://huggingface.co/datasets?search=EMG-Dataset) — Hugging Face · 19,767 files · 675,553 bytes · revision `6c8f265` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0007) | | [Ann000/xr-emg-force-interface ↗](https://huggingface.co/datasets?search=xr-emg-force-interface) — Hugging Face · 1,804 files · 13,976,490,272 bytes · revision `b4f2ed0` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0008) | | [ApocalyVec/olive-physio ↗](https://huggingface.co/datasets?search=olive-physio) — Hugging Face · 5 files · 3,188,786,217 bytes · revision `2dbbeb1` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0009) | | [Aurelianous/eeg-audio ↗](https://huggingface.co/datasets?search=eeg-audio) — Hugging Face · 194 files · 2,456,586,217 bytes · revision `5f26b50` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0010) | | [Bhuvii19/EEGFakeBench ↗](https://huggingface.co/datasets?search=EEGFakeBench) — Hugging Face · 1,064 files · 39,885,528,289 bytes · revision `5fbff07` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0011) | | [BiscuitBobby/eeg_cvpr ↗](https://huggingface.co/datasets?search=eeg_cvpr) — Hugging Face · 2 files · 10,975,593,526 bytes · revision `cc85cfe` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0013) | | [BrainPilot-Bench/Tasks-Data-Public ↗](https://huggingface.co/datasets?search=Tasks-Data-Public) — Hugging Face · 898 files · 37,946,341,174 bytes · revision `70efdcf` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0014) | | [ChenglinLiuChris/eeg ↗](https://huggingface.co/datasets?search=eeg) — Hugging Face · 2 files · 2,844,537,640 bytes · revision `074a02b` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0016) | | [Dadaism6/Omni-iEEG-Channel-feature ↗](https://huggingface.co/datasets?search=Omni-iEEG-Channel-feature) — Hugging Face · 619 files · 144,118,982,579 bytes · revision `47c43d1` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0017) | | [Dadaism6/Omni-iEEG-Raw-Event-EDF ↗](https://huggingface.co/datasets?search=Omni-iEEG-Raw-Event-EDF) — Hugging Face · 55 files · 8,536,463,816 bytes · revision `3ebc708` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0018) | | [Emulated-Inc/drowsy-driving-reaction-time-eeg ↗](https://huggingface.co/datasets?search=drowsy-driving-reaction-time-eeg) — Hugging Face · 103 files · 9,820,339,528 bytes · revision `fc6b8fb` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0019) | | [Emulated-Inc/meg-story-phoneme-decoding ↗](https://huggingface.co/datasets?search=meg-story-phoneme-decoding) — Hugging Face · 7 files · 7,337,303,936 bytes · revision `e08b34c` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0020) | | [Emulated-Inc/mi-eeg-unseen-subjects ↗](https://huggingface.co/datasets?search=mi-eeg-unseen-subjects) — Hugging Face · 10 files · 1,333,496,221 bytes · revision `681e5c0` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names Stieger longitudinal BCI · one-person pilot as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0021) | | [Emulated-Inc/sci-emg-gesture-new-person ↗](https://huggingface.co/datasets?search=sci-emg-gesture-new-person) — Hugging Face · 14 files · 7,070,691,235 bytes · revision `5dec83f` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0022) | | [Emulated-Inc/sci-ieeg-clip-triage-new-centre ↗](https://huggingface.co/datasets?search=sci-ieeg-clip-triage-new-centre) — Hugging Face · 20 files · 1,549,429,060 bytes · revision `014fc29` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0023) | | [Emulated-Inc/seeg-speech-spectrogram-reconstruction ↗](https://huggingface.co/datasets?search=seeg-speech-spectrogram-reconstruction) — Hugging Face · 4 files · 2,794,940,875 bytes · revision `79179af` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0024) | | [Emulated-Inc/silent-speech-emg-decoding ↗](https://huggingface.co/datasets?search=silent-speech-emg-decoding) — Hugging Face · 6 files · 5,432,414,411 bytes · revision `dc7f3b7` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0025) | | [Emulated-Inc/ssvep-calibration-free-short-window ↗](https://huggingface.co/datasets?search=ssvep-calibration-free-short-window) — Hugging Face · 37 files · 2,549,292,653 bytes · revision `cc55ae8` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names BETA as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0026) | | [Emulated-Inc/things-eeg-zero-shot-concepts ↗](https://huggingface.co/datasets?search=things-eeg-zero-shot-concepts) — Hugging Face · 5 files · 2,895,140,376 bytes · revision `49a19b0` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0027) | | [Felipe717/wbcic-2c-ica ↗](https://huggingface.co/datasets?search=wbcic-2c-ica) — Hugging Face · 3 files · 736,241,643 bytes · revision `bb6603b` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names WBCIC-SHU motor imagery dataset as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0028) | | [Forainest789/NATVIEW_EEGFMRI_processed ↗](https://huggingface.co/datasets?search=NATVIEW_EEGFMRI_processed) — Hugging Face · 24,576 files · 31,134,649,016 bytes · revision `8acf40d` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0029) | | [Fudan-fMRI/CineBrain ↗](https://huggingface.co/datasets?search=CineBrain) — Hugging Face · 19 files · 140,971,724,389 bytes · revision `bb3afbd` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0030) | | [Fudan-fMRI/EEG-Graspbot ↗](https://huggingface.co/datasets?search=EEG-Graspbot) — Hugging Face · 97 files · 15,236,892,075 bytes · revision `d7d9c1f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0031) | | [Haitao999/things-eeg ↗](https://huggingface.co/datasets?search=things-eeg) — Hugging Face · 33,519 files · 27,830,568,428 bytes · revision `944f0d5` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0032) | | [HeliumHcc/EEG_raw_datasets ↗](https://huggingface.co/datasets?search=EEG_raw_datasets) — Hugging Face · 14 files · 130,957,002,576 bytes · revision `7976bc2` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request.; Matches BCI Competition IV data sets in archive names ([terms page captured 19 October 2026 ↗](https://www.bbci.de/competition/iv/)).; Matches FACED (Finer-grained Affective Computing EEG Dataset) in archive names ([terms page captured 19 October 2026 ↗](https://www.nature.com/articles/s41597-023-02650-w)).; Matches TUH EEG Corpus in archive names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0033) | | [HeliumHcc/preprocessed_EEG ↗](https://huggingface.co/datasets?search=preprocessed_EEG) — Hugging Face · 12 files · 40,910,679,856 bytes · revision `dcaf69e` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0034) | | [Ishah8840/emg-ego-multiview-hand-pose ↗](https://huggingface.co/datasets?search=emg-ego-multiview-hand-pose) — Hugging Face · 124 files · 6,699,433,595 bytes · revision `0f2abdb` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0035) | | [JLB-JLB/seizure_eeg_dev ↗](https://huggingface.co/datasets?search=seizure_eeg_dev) — Hugging Face · 9 files · 3,418,836,109 bytes · revision `f717a86` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Matches TUH EEG Corpus in example counts, label names, and split names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0036) | | [JLB-JLB/seizure_eeg_eval ↗](https://huggingface.co/datasets?search=seizure_eeg_eval) — Hugging Face · 9 files · 3,418,836,110 bytes · revision `208377f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0037) | | [JLB-JLB/seizure_eeg_greyscale_224x224_6secWindow ↗](https://huggingface.co/datasets?search=seizure_eeg_greyscale_224x224_6secWindow) — Hugging Face · 82 files · 39,216,540,324 bytes · revision `08da921` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Matches TUH EEG Corpus in the source its uploader’s code names, label names, and split names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0038) | | [JLB-JLB/seizure_eeg_greyscale_224x224_6secWindow_adjusted ↗](https://huggingface.co/datasets?search=seizure_eeg_greyscale_224x224_6secWindow_adjusted) — Hugging Face · 33 files · 14,909,837,243 bytes · revision `8909a00` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0039) | | [JLB-JLB/seizure_eeg_iirFilter_greyscale_224x224_6secWindow ↗](https://huggingface.co/datasets?search=seizure_eeg_iirFilter_greyscale_224x224_6secWindow) — Hugging Face · 83 files · 39,464,496,948 bytes · revision `f0a39e3` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0040) | | [JLB-JLB/seizure_eeg_iirFilter_greyscale_224x224_6secWindow_adjusted ↗](https://huggingface.co/datasets?search=seizure_eeg_iirFilter_greyscale_224x224_6secWindow_adjusted) — Hugging Face · 34 files · 14,973,762,990 bytes · revision `d698090` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Matches TUH EEG Corpus in example counts, label names, and split names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0041) | | [JLB-JLB/seizure_eeg_train ↗](https://huggingface.co/datasets?search=seizure_eeg_train) — Hugging Face · 50 files · 24,165,940,730 bytes · revision `e83ff38` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Matches TUH EEG Corpus in example counts, label names, and split names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0042) | | [James-ChenSD/rawData_EEG ↗](https://huggingface.co/datasets?search=rawData_EEG) — Hugging Face · 10 files · 28,983,980,808 bytes · revision `bc7d518` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0043) | | [JayF14/chb-mit-preprocessed-eeg ↗](https://huggingface.co/datasets?search=chb-mit-preprocessed-eeg) — Hugging Face · 26 files · 4,058,065,322 bytes · revision `8ac9ea1` | — | `odc-by` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0044) | | [Jingrui77/EMGVLA-Dataset ↗](https://huggingface.co/datasets?search=EMGVLA-Dataset) — Hugging Face · 36,738 files · 33,967,013,309 bytes · revision `f62b8f6` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0046) | | [JuniorThap/EEG-relaxation-concentration ↗](https://huggingface.co/datasets?search=EEG-relaxation-concentration) — Hugging Face · 5 files · 214,490,363 bytes · revision `2fd124f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0047) | | [KernelCo/holohub_bci_visualization ↗](https://huggingface.co/datasets?search=holohub_bci_visualization) — Hugging Face · 92 files · 74,199,500,093 bytes · revision `fe75bc1` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0048) | | [Krishan0698/BCI ↗](https://huggingface.co/datasets?search=BCI) — Hugging Face · 2 files · 252,214,492 bytes · revision `d64ce87` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0057) | | [Kukedlc/SpanishBCBL ↗](https://huggingface.co/datasets?search=SpanishBCBL) — Hugging Face · 731 files · 280,382,552,015 bytes · revision `6a3118a` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0058) | | [Lance1573/L-Mind ↗](https://huggingface.co/datasets?search=L-Mind) — Hugging Face · 12,012 files · 2,422,510,990 bytes · revision `31fc106` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0060) | | [LeBabyOx/EEGParquet ↗](https://huggingface.co/datasets?search=EEGParquet) — Hugging Face · 689 files · 3,875,553,348 bytes · revision `76f5f55` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0061) | | [LeRobot-worldwide-hackathon/7-LeCyborg-so100_emg_sensor ↗](https://huggingface.co/datasets?search=7-LeCyborg-so100_emg_sensor) — Hugging Face · 66 files · 116,220,150 bytes · revision `4e9c919` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0062) | | [LidongYang/BrainFLORA ↗](https://huggingface.co/datasets?search=BrainFLORA) — Hugging Face · 31,030 files · 101,004,899,421 bytes · revision `210051a` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches THINGS-data (THINGS-MEG and THINGS-fMRI) in the source its uploader’s code names, example counts, and participant count. That source’s terms page gives `cc0-1.0`; this upload’s licence tag is `apache-2.0` ([terms page captured 19 October 2026 ↗](https://github.com/OpenNeuroDatasets/ds004212/blob/136f5297b07d930028fb5a4a16e649efa728aaf5/dataset_description.json)).; Matches THINGS object image database in the source its uploader’s code names and example counts ([terms page captured 19 October 2026 ↗](https://osf.io/download/ahg2k/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0063) | | [LidongYang/EEG_Image_decode ↗](https://huggingface.co/datasets?search=EEG_Image_decode) — Hugging Face · 71 files · 159,575,450,629 bytes · revision `dbe34bb` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0064) | | [LiyuanHan/bcijelly-invasive ↗](https://huggingface.co/datasets?search=bcijelly-invasive) — Hugging Face · 849 files · 5,274,689,570 bytes · revision `c38668c` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0065) | | [Lo6yu/egocentric_dataset ↗](https://huggingface.co/datasets?search=egocentric_dataset) — Hugging Face · 6,127 files · 30,118,622,892 bytes · revision `a318ea3` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0066) | | MER-PS/MER-PS-public-leaderboard-evaluation-data — Hugging Face · 27 files · 4,050,355,685 bytes · revision `28acc44` | — | `cc-by-nc-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: access is granted automatically once a user accepts the conditions. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0067) | | MER-PS/MER-PS-trainval — Hugging Face · 3 files · 22,317,841,274 bytes · revision `7762273` | — | `cc-by-nc-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: access is granted automatically once a user accepts the conditions. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0068) | | [MElHuseyni/EEG_data ↗](https://huggingface.co/datasets?search=EEG_data) — Hugging Face · 4 files · 4,810,895 bytes · revision `f6031a0` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0069) | | [MaRez10/EEG_Pain_Perception ↗](https://huggingface.co/datasets?search=EEG_Pain_Perception) — Hugging Face · 3 files · 12,738,959,235 bytes · revision `36bbfd8` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: access is granted automatically once a user accepts the conditions. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0070) | | [MarksonChen/flare ↗](https://huggingface.co/datasets?search=flare) — Hugging Face · 2,125 files · 247,359,935,324 bytes · revision `a280589` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names MC_RTT (Neural Latents Benchmark) as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0071) | | [McGregorW/EEGDatasets ↗](https://huggingface.co/datasets?search=EEGDatasets) — Hugging Face · 29 files · 228,941,021,425 bytes · revision `a4572e6` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request.; Matches CHB-MIT Scalp EEG Database (PhysioNet chbmit 1.0.0) in archive names. That source’s terms page gives `odc-by-1.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/chbmit/1.0.0/)).; Matches EEGMAT in archive names. That source’s terms page gives `odc-by-1.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmat/1.0.0/)).; Matches physionet-eegmmidb-1.0.0 in archive names. That source’s terms page gives `odc-by-1.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)).; Matches Haaglanden Medisch Centrum sleep staging database (PhysioNet hmc-sleep-staging 1.1) in archive names. That source’s terms page gives `cc-by-4.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/hmc-sleep-staging/1.1/)).; Matches Siena Scalp EEG Database (PhysioNet siena-scalp-eeg 1.0.0) in archive names. That source’s terms page gives `cc-by-4.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/siena-scalp-eeg/1.0.0/)).; Matches Sleep-EDF Database Expanded (PhysioNet sleep-edfx 1.0.0) in archive names. That source’s terms page gives `odc-by-1.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/sleep-edfx/1.0.0/)).; Matches SEED family in archive names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://bcmi.sjtu.edu.cn/home/seed/downloads.html)).; Matches Sleep Heart Health Study (SHHS, NSRR) in archive names. That source is available on request; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://sleepdata.org/data/requests/shhs/start)).; Matches TUH EEG Corpus in archive names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0072) | | [McGregorW/EEGDatasetsH5 ↗](https://huggingface.co/datasets?search=EEGDatasetsH5) — Hugging Face · 448 files · 279,038,350,289 bytes · revision `42a2b23` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request.; Matches CAUEEG (Chung-Ang University Hospital EEG dataset) in folder names. That source is available on request; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://github.com/ipis-mjkim/caueeg-dataset/blob/033a2e5ddc837ae0022f9add07aef8e10bcd4ad1/README.md)).; Matches CHB-MIT Scalp EEG Database (PhysioNet chbmit 1.0.0) in folder names. That source’s terms page gives `odc-by-1.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/chbmit/1.0.0/)).; Matches Siena Scalp EEG Database (PhysioNet siena-scalp-eeg 1.0.0) in folder names. That source’s terms page gives `cc-by-4.0`; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://physionet.org/content/siena-scalp-eeg/1.0.0/)).; Matches SEED family in folder names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://bcmi.sjtu.edu.cn/home/seed/downloads.html)).; Matches Sleep Heart Health Study (SHHS, NSRR) in folder names. That source is available on request; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://sleepdata.org/data/requests/shhs/start)).; Matches TUH EEG Corpus in folder names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0073) | | [McGregorW/EEGDatasets_raw ↗](https://huggingface.co/datasets?search=EEGDatasets_raw) — Hugging Face · 3 files · 32,350,523,184 bytes · revision `b1391e1` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0074) | | [MigoXV/eeg-bme-emotion-parquet ↗](https://huggingface.co/datasets?search=eeg-bme-emotion-parquet) — Hugging Face · 5 files · 994,152,644 bytes · revision `1921498` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0075) | | [MrC4t/LeCyborg-so100_emg_sensor ↗](https://huggingface.co/datasets?search=LeCyborg-so100_emg_sensor) — Hugging Face · 66 files · 116,220,265 bytes · revision `e546479` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0076) | | [NOttheol/EEG-Talha-Alakus-Gonen-Turkoglu ↗](https://huggingface.co/datasets?search=EEG-Talha-Alakus-Gonen-Turkoglu) — Hugging Face · 221 files · 633,592,821 bytes · revision `9cf5544` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0077) | | [Nerlnet/EEG_Eye_State_Classification ↗](https://huggingface.co/datasets?search=EEG_Eye_State_Classification) — Hugging Face · 4 files · 3,133,497 bytes · revision `97c351d` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0078) | | [Nerlnet/EEG_Eye_State_Classification_table_freq_window ↗](https://huggingface.co/datasets?search=EEG_Eye_State_Classification_table_freq_window) — Hugging Face · 3 files · 5,993,741 bytes · revision `a16a633` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0079) | | [NeuroBench/thor_eeg_mi ↗](https://huggingface.co/datasets?search=thor_eeg_mi) — Hugging Face · 5 files · 1,125,003,592 bytes · revision `dd0388b` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0080) | | NeuroMM/NeuroMM-2026 — Hugging Face · 23 files · 15,835,588,605 bytes · revision `dcd61c4` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0081) | | [NeuroTec/SWEC_iEEG_Dataset ↗](https://huggingface.co/datasets?search=SWEC_iEEG_Dataset) — Hugging Face · 694 files · 4,797,015,502,765 bytes · revision `584e9d2` | — | `cdla-permissive-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0082) | | Omni-iEEG/Omni-iEEG — Hugging Face · 4,277 files · 159,247,652,055 bytes · revision `73b9c51` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0083) | | [PaoloMaccallini/EEG ↗](https://huggingface.co/datasets?search=EEG) — Hugging Face · 443 files · 2,025,587,255 bytes · revision `617e1dc` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0084) | | [Papyson/Music_Stimulated_EEG ↗](https://huggingface.co/datasets?search=Music_Stimulated_EEG) — Hugging Face · 3 files · 869,051,650 bytes · revision `2605b14` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0085) | | [PromiseZ5Q2SQ/EEG-Hallucination ↗](https://huggingface.co/datasets?search=EEG-Hallucination) — Hugging Face · 157 files · 26,751,300,311 bytes · revision `a002c6e` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0086) | | [PromiseZ5Q2SQ/EEG-ImageNet ↗](https://huggingface.co/datasets?search=EEG-ImageNet) — Hugging Face · 3 files · 14,351,613,293 bytes · revision `2594358` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0087) | | [Quoron/EEG-semantic-text-relevance ↗](https://huggingface.co/datasets?search=EEG-semantic-text-relevance) — Hugging Face · 27 files · 11,928,213,255 bytes · revision `96b0b03` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0088) | | [Rodr16020/eeg-words-dataset ↗](https://huggingface.co/datasets?search=eeg-words-dataset) — Hugging Face · 4 files · 29,936,552 bytes · revision `29f1732` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0089) | | [RuthvikBandari/neural-compass-bci ↗](https://huggingface.co/datasets?search=neural-compass-bci) — Hugging Face · 167 files · 6,120,316,023 bytes · revision `0f91481` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0090) | | [SamLegion888/SpanishBCBL ↗](https://huggingface.co/datasets?search=SpanishBCBL) — Hugging Face · 731 files · 280,382,552,015 bytes · revision `21e7071` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0091) | | [Samuelsantos777/psg-audio-v3-unofficial-mirror ↗](https://huggingface.co/datasets?search=psg-audio-v3-unofficial-mirror) — Hugging Face · 2,022 files · 984,855,341,189 bytes · revision `cb22565` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0092) | | [Scan2s/ninapro-db5-s2s-certified ↗](https://huggingface.co/datasets?search=ninapro-db5-s2s-certified) — Hugging Face · 3 files · 1,101,828 bytes · revision `0040e87` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0095) | | [SciAILab/OSWorld-Science-data ↗](https://huggingface.co/datasets?search=OSWorld-Science-data) — Hugging Face · 2,759 files · 239,043,791,977 bytes · revision `9b8de9a` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0096) | | [StevenLe456/psychiatric-eeg ↗](https://huggingface.co/datasets?search=psychiatric-eeg) — Hugging Face · 2 files · 10,539,714 bytes · revision `a220fb9` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0097) | | [TKQXX/THINGS-EEG-MEG_CVPR2026 ↗](https://huggingface.co/datasets?search=THINGS-EEG-MEG_CVPR2026) — Hugging Face · 85,727 files · 42,034,207,638 bytes · revision `3a57e9c` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0098) | | [Tanisha0709/EEGBCI ↗](https://huggingface.co/datasets?search=EEGBCI) — Hugging Face · 51 files · 1,044,968,369 bytes · revision `8c57f05` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0099) | | [Terrycz/ECOG-ds005574 ↗](https://huggingface.co/datasets?search=ECOG-ds005574) — Hugging Face · 154 files · 15,507,112,315 bytes · revision `d844dde` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0100) | | [Terrycz/SEEG-singlewordDutch ↗](https://huggingface.co/datasets?search=SEEG-singlewordDutch) — Hugging Face · 125 files · 5,312,314,072 bytes · revision `2e1e514` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0101) | | [Twu31/cap32-mi-eeg ↗](https://huggingface.co/datasets?search=cap32-mi-eeg) — Hugging Face · 26 files · 69,901,595 bytes · revision `663cd80` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0102) | | [Twu31/cap32-ssvep ↗](https://huggingface.co/datasets?search=cap32-ssvep) — Hugging Face · 11 files · 4,160,406 bytes · revision `8096d4b` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0103) | | [Twu31/ssvep-jfpm-12target ↗](https://huggingface.co/datasets?search=ssvep-jfpm-12target) — Hugging Face · 10 files · 3,693,868 bytes · revision `1360ca5` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0104) | | [USuCgex0122e/demo_dataset_seed20260726 ↗](https://huggingface.co/datasets?search=demo_dataset_seed20260726) — Hugging Face · 147 files · 41,422,370,986 bytes · revision `41602cf` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Its card names DEAP as its source.; Its card names TUH EEG Corpus as its source.; Matches DEAP in channel count, folder names, and participant count. That source is released under a signed agreement; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://www.eecs.qmul.ac.uk/mmv/datasets/deap/download.html)).; Matches MODMA (Multi-modal Open Dataset for Mental-disorder Analysis) in channel count, folder names, and participant count. That source is released under a signed agreement; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://modma.lzu.edu.cn/data/application/)).; Matches CHB-MIT Scalp EEG Database (PhysioNet chbmit 1.0.0) in folder names and participant count. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/chbmit/1.0.0/)).; Matches EEGMAT in folder names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmat/1.0.0/)).; Matches physionet-eegmmidb-1.0.0 in folder names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)).; Matches Siena Scalp EEG Database (PhysioNet siena-scalp-eeg 1.0.0) in folder names and participant count ([terms page captured 19 October 2026 ↗](https://physionet.org/content/siena-scalp-eeg/1.0.0/)).; Matches Sleep-EDF Database Expanded (PhysioNet sleep-edfx 1.0.0) in folder names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/sleep-edfx/1.0.0/)).; Matches SEED family in folder names and participant count. That source is released under a signed agreement; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://bcmi.sjtu.edu.cn/home/seed/downloads.html)).; Matches TUH EEG Corpus in folder names and label names. That source is released under a signed agreement; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0105) | | [Ugo96/MMI-EEGSR ↗](https://huggingface.co/datasets?search=MMI-EEGSR) — Hugging Face · 3 files · 1,425,353,838 bytes · revision `897e0a0` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches physionet-eegmmidb-1.0.0 in the source its uploader’s code names and dataset name. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0106) | | [Voxel51/Egocentric-EMG-Force ↗](https://huggingface.co/datasets?search=Egocentric-EMG-Force) — Hugging Face · 14 files · 8,433,938,159 bytes · revision `6882191` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0107) | | [WindySprint/EESM19-Processed ↗](https://huggingface.co/datasets?search=EESM19-Processed) — Hugging Face · 4 files · 88,599,844,434 bytes · revision `a3b7abb` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names EESM19 scalp subset as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0108) | | [WindySprint/EESM23-Processed ↗](https://huggingface.co/datasets?search=EESM23-Processed) — Hugging Face · 4 files · 5,601,567,078 bytes · revision `c116776` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names EESM23 · Ear-EEG Sleep Monitoring 2023 as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0109) | | [WymanYYY/Sleep-EDF-V2 ↗](https://huggingface.co/datasets?search=Sleep-EDF-V2) — Hugging Face · 6 files · 20,896,170,531 bytes · revision `63af957` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0110) | | [Y-Research-Group/NeuroSonic ↗](https://huggingface.co/datasets?search=NeuroSonic) — Hugging Face · 21 files · 72,632,019,873 bytes · revision `ba1498d` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0111) | | [Yvhang009/eeg-fmri-ds006040-derived ↗](https://huggingface.co/datasets?search=eeg-fmri-ds006040-derived) — Hugging Face · 1,199 files · 14,253,807,366 bytes · revision `cfa605c` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0112) | | [Yvhang009/eeg-fmri-natview-derived ↗](https://huggingface.co/datasets?search=eeg-fmri-natview-derived) — Hugging Face · 1,235 files · 3,580,011,514 bytes · revision `211ea56` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0113) | | [Zachary1150/EESM17-Processed ↗](https://huggingface.co/datasets?search=EESM17-Processed) — Hugging Face · 12 files · 14,280,315,243 bytes · revision `269028b` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0114) | | [Zachary1150/EESM19-Processed ↗](https://huggingface.co/datasets?search=EESM19-Processed) — Hugging Face · 6 files · 124,077,444,322 bytes · revision `fb04501` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names EESM19 scalp subset as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0115) | | [Zachary1150/EESM23-Processed ↗](https://huggingface.co/datasets?search=EESM23-Processed) — Hugging Face · 11 files · 20,170,017,458 bytes · revision `33cf229` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names EESM23 · Ear-EEG Sleep Monitoring 2023 as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0116) | | [Zachary1150/EventAAD-Processed ↗](https://huggingface.co/datasets?search=EventAAD-Processed) — Hugging Face · 4 files · 6,225,968,056 bytes · revision `4998746` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0117) | | [Zachary1150/Motor-Processed ↗](https://huggingface.co/datasets?search=Motor-Processed) — Hugging Face · 4 files · 1,740,145,191 bytes · revision `4e654b8` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0118) | | [Zachary1150/ear-eeg-fm-benchmark-data ↗](https://huggingface.co/datasets?search=ear-eeg-fm-benchmark-data) — Hugging Face · 19 files · 33,956,076,983 bytes · revision `39a7821` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing.; Its card names EESM23 · Ear-EEG Sleep Monitoring 2023 as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0119) | | [Zachary1150/ear-eeg-motor-raw ↗](https://huggingface.co/datasets?search=ear-eeg-motor-raw) — Hugging Face · 8 files · 6,859,144,767 bytes · revision `cc1c273` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0120) | | [aayushi83730709/emg-gestures-task1 ↗](https://huggingface.co/datasets?search=emg-gestures-task1) — Hugging Face · 438 files · 3,755,871,250 bytes · revision `cc89e01` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0121) | | [abachu2005/mes-eeg-processed ↗](https://huggingface.co/datasets?search=mes-eeg-processed) — Hugging Face · 121 files · 148,041,942 bytes · revision `e9e1525` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0122) | | [abdo25/EEG_records ↗](https://huggingface.co/datasets?search=EEG_records) — Hugging Face · 2 files · 28,232,412 bytes · revision `04981a9` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0123) | | [alexeykashevnik/EEGMeditation ↗](https://huggingface.co/datasets?search=EEGMeditation) — Hugging Face · 549 files · 312,534,424,980 bytes · revision `a44324f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: access is granted automatically once a user accepts the conditions. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0124) | | [alibabasglab/KUL-mix ↗](https://huggingface.co/datasets?search=KUL-mix) — Hugging Face · 3 files · 3,089,025,488 bytes · revision `b8aab92` | — | `cc-by-nc-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0125) | | [alif-munim/eeg-filtered ↗](https://huggingface.co/datasets?search=eeg-filtered) — Hugging Face · 4 files · 571,769,037 bytes · revision `b1fd9f7` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0126) | | [amongglue/amber ↗](https://huggingface.co/datasets?search=amber) — Hugging Face · 11,727 files · 109,323,974,936 bytes · revision `2ca0bbc` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches BNCI Horizon 2020 data sets catalogue in the source its card names ([terms page captured 19 October 2026 ↗](https://bnci-horizon-2020.eu/database/data-sets)).; Matches Healthy Brain Network (HBN, Child Mind Institute) in the source its card names ([terms page captured 19 October 2026 ↗](https://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/FAQ.html)).; Matches CHB-MIT Scalp EEG Database (PhysioNet chbmit 1.0.0) in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/chbmit/1.0.0/)).; Matches EEGMAT in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmat/1.0.0/)).; Matches physionet-eegmmidb-1.0.0 in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)).; Matches ERP-based Brain-Computer Interface recordings (PhysioNet erpbci 1.0.0) in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/erpbci/1.0.0/)).; Matches Haaglanden Medisch Centrum sleep staging database (PhysioNet hmc-sleep-staging 1.1) in the source its card names. That source’s terms page gives `cc-by-4.0`; its card gives `odc-by-1.0` for this source ([terms page captured 19 October 2026 ↗](https://physionet.org/content/hmc-sleep-staging/1.1/)).; Matches LTRSVP · EEG Signals from an RSVP Task in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/ltrsvp/1.0.0/)).; Matches Sleep Heart Health Study PSG Database (PhysioNet shhpsgdb 1.0.0) in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/shhpsgdb/1.0.0/)).; Matches Siena Scalp EEG Database (PhysioNet siena-scalp-eeg 1.0.0) in the source its card names. That source’s terms page gives `cc-by-4.0`; its card gives `odc-by-1.0` for this source ([terms page captured 19 October 2026 ↗](https://physionet.org/content/siena-scalp-eeg/1.0.0/)).; Matches Sleep-EDF Database Expanded (PhysioNet sleep-edfx 1.0.0) in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/sleep-edfx/1.0.0/)).; Matches St. Vincent's University Hospital / University College Dublin Sleep Apnea Database (PhysioNet ucddb 1.0.0) in the source its card names ([terms page captured 19 October 2026 ↗](https://physionet.org/content/ucddb/1.0.0/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0127) | | [aparker03/eeg-csv ↗](https://huggingface.co/datasets?search=eeg-csv) — Hugging Face · 214 files · 1,618,761,261 bytes · revision `6b2af64` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0128) | | [argaaw/aved_eeg ↗](https://huggingface.co/datasets?search=aved_eeg) — Hugging Face · 2 files · 1,593,967,389 bytes · revision `012d88b` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0129) | | [argaaw/dtu_eeg ↗](https://huggingface.co/datasets?search=dtu_eeg) — Hugging Face · 2 files · 1,769,888,541 bytes · revision `d147eac` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0130) | | [argaaw/dtu_eeg_more_preproc ↗](https://huggingface.co/datasets?search=dtu_eeg_more_preproc) — Hugging Face · 2 files · 885,152,541 bytes · revision `3ca84d8` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0131) | | [as674262040/pub_eeg ↗](https://huggingface.co/datasets?search=pub_eeg) — Hugging Face · 1,706 files · 57,125,727,523 bytes · revision `5dedd2a` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0132) | | [aspire-osu/maestro-eeg-dataset ↗](https://huggingface.co/datasets?search=maestro-eeg-dataset) — Hugging Face · 9,055 files · 41,870,569,843 bytes · revision `de5773a` | — | `cc-by-nc-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: access is granted automatically once a user accepts the conditions. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0133) | | [atgarcia/EMGSoundFull ↗](https://huggingface.co/datasets?search=EMGSoundFull) — Hugging Face · 28 files · 4,788,839,046 bytes · revision `d91b344` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0134) | | [atgarcia/EMGSoundTest ↗](https://huggingface.co/datasets?search=EMGSoundTest) — Hugging Face · 8 files · 1,587,698,562 bytes · revision `359450d` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0135) | | [basematrix/ego-binocular-v1 ↗](https://huggingface.co/datasets?search=ego-binocular-v1) — Hugging Face · 55 files · 7,958,424,208 bytes · revision `d6c7a48` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0140) | | [bcbl190626/SpanishBCBL ↗](https://huggingface.co/datasets?search=SpanishBCBL) — Hugging Face · 731 files · 280,382,552,015 bytes · revision `88f9096` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0141) | | [brain-bzh/eegmat-prepro ↗](https://huggingface.co/datasets?search=eegmat-prepro) — Hugging Face · 5 files · 274,699,801 bytes · revision `754f567` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing.; Its card names EEGMAT as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0142) | | [brain-bzh/reve-dataset ↗](https://huggingface.co/datasets?search=reve-dataset) — Hugging Face · 1,079 files · 4,849,604,127,659 bytes · revision `f07140e` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0143) | | [braindecode/arithmetic_zyma2019 ↗](https://huggingface.co/datasets?search=arithmetic_zyma2019) — Hugging Face · 510 files · 122,373,830 bytes · revision `8359206` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0144) | | [braindecode/bcic2020-3 ↗](https://huggingface.co/datasets?search=bcic2020-3) — Hugging Face · 351 files · 853,715,845 bytes · revision `8e2adde` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0145) | | [braindecode/bcic2a ↗](https://huggingface.co/datasets?search=bcic2a) — Hugging Face · 762 files · 343,324,140 bytes · revision `3585206` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0146) | | [braindecode/chbmit ↗](https://huggingface.co/datasets?search=chbmit) — Hugging Face · 4,793 files · 4,711,410,299 bytes · revision `3c7fc55` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0147) | | [braindecode/faced ↗](https://huggingface.co/datasets?search=faced) — Hugging Face · 990 files · 3,716,389,982 bytes · revision `18c7335` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0151) | | [braindecode/isruc-sleep ↗](https://huggingface.co/datasets?search=isruc-sleep) — Hugging Face · 1,392 files · 13,105,355,460 bytes · revision `beda681` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0152) | | [braindecode/mdd_mumtaz2016 ↗](https://huggingface.co/datasets?search=mdd_mumtaz2016) — Hugging Face · 1,273 files · 1,046,316,816 bytes · revision `53cd9de` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0153) | | [braindecode/physionet ↗](https://huggingface.co/datasets?search=physionet) — Hugging Face · 4,584 files · 1,450,253,680 bytes · revision `3c9cf8b` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0154) | | [cagataydev/emotiv-ecot ↗](https://huggingface.co/datasets?search=emotiv-ecot) — Hugging Face · 7 files · 228,087 bytes · revision `f731954` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0155) | | [charantejarangi123/EEG_dataset ↗](https://huggingface.co/datasets?search=EEG_dataset) — Hugging Face · 2 files · 474,228,379 bytes · revision `f47e6fc` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0156) | | [conorhassan/fast-autoregressive-inference-eeg ↗](https://huggingface.co/datasets?search=fast-autoregressive-inference-eeg) — Hugging Face · 10,052 files · 2,666,627,427 bytes · revision `4a5921c` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0157) | | [ctorre/self-calibrating-bci ↗](https://huggingface.co/datasets?search=self-calibrating-bci) — Hugging Face · 10 files · 40,826,464 bytes · revision `dadba68` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0158) | | [dancinlab/anima-eeg-consciousness ↗](https://huggingface.co/datasets?search=anima-eeg-consciousness) — Hugging Face · 10 files · 10,785,549 bytes · revision `3462d99` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0159) | | [deremustapha/FlexAdapt_EMG_Dataset ↗](https://huggingface.co/datasets?search=FlexAdapt_EMG_Dataset) — Hugging Face · 1,114 files · 87,076,006 bytes · revision `65a2716` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0161) | | [dhioewhd/things-eeg ↗](https://huggingface.co/datasets?search=things-eeg) — Hugging Face · 33,519 files · 27,830,568,428 bytes · revision `041e68f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0162) | | [divyanshmaurya1/BCI_Data_new ↗](https://huggingface.co/datasets?search=BCI_Data_new) — Hugging Face · 102 files · 272,506,963 bytes · revision `fc1d17a` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Its card names Cho2017 as its source.; Matches Cho et al. 2017 motor imagery EEG (GigaDB 100295) in the source its card names and file names. That source’s terms page gives `cc0-1.0`; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://doi.org/10.5524/100295)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0163) | | [divyanshmaurya1/projectcerebro-eeg ↗](https://huggingface.co/datasets?search=projectcerebro-eeg) — Hugging Face · 4 files · 165,398,589 bytes · revision `cbfca4b` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0164) | | [divyanshmaurya1/projectcerebro-eeg-parquet ↗](https://huggingface.co/datasets?search=projectcerebro-eeg-parquet) — Hugging Face · 2 files · 165,388,570 bytes · revision `e6fc153` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0165) | | [eZWALT/price-of-attention ↗](https://huggingface.co/datasets?search=price-of-attention) — Hugging Face · 54 files · 4,288,935 bytes · revision `6fdd999` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0166) | | [edessa/EMG ↗](https://huggingface.co/datasets?search=EMG) — Hugging Face · 703 files · 675,745,479,043 bytes · revision `66b9aa5` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0167) | | [fabricioslv/omnimind-neuro-henry-obleser-meg-amfm ↗](https://huggingface.co/datasets?search=omnimind-neuro-henry-obleser-meg-amfm) — Hugging Face · 3 files · 3,465,177,068 bytes · revision `869163d` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0168) | | [fenjoyoung/things-eeg ↗](https://huggingface.co/datasets?search=things-eeg) — Hugging Face · 33,519 files · 27,830,568,428 bytes · revision `58ee885` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0169) | | [fracapuano/eeg2meg-large-tokenized ↗](https://huggingface.co/datasets?search=eeg2meg-large-tokenized) — Hugging Face · 24 files · 5,263,856,517 bytes · revision `4b6a8ba` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0170) | | [fracapuano/eeg2meg-medium ↗](https://huggingface.co/datasets?search=eeg2meg-medium) — Hugging Face · 6 files · 1,852,370,434 bytes · revision `c07a283` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0171) | | [fracapuano/eeg2meg-medium-tokenized ↗](https://huggingface.co/datasets?search=eeg2meg-medium-tokenized) — Hugging Face · 7 files · 1,355,157,621 bytes · revision `8958aec` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0172) | | [fracapuano/eeg2meg-small-tokenized ↗](https://huggingface.co/datasets?search=eeg2meg-small-tokenized) — Hugging Face · 3 files · 587,554 bytes · revision `f8304fe` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0173) | | [fracapuano/eeg2meg-tiny ↗](https://huggingface.co/datasets?search=eeg2meg-tiny) — Hugging Face · 3 files · 124,341,505 bytes · revision `de0d242` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0174) | | [frthjf/spe1-zhao2026-benchmark ↗](https://huggingface.co/datasets?search=spe1-zhao2026-benchmark) — Hugging Face · 458 files · 157,090,735,679 bytes · revision `697ff26` | — | `cc-by-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0175) | | [gabrycina/eeg2meg-medium ↗](https://huggingface.co/datasets?search=eeg2meg-medium) — Hugging Face · 7 files · 1,617,384,905 bytes · revision `142c057` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0176) | | [gabrycina/eeg2meg-tiny ↗](https://huggingface.co/datasets?search=eeg2meg-tiny) — Hugging Face · 5 files · 91,339,453 bytes · revision `3a4193f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0177) | | [gasparyanartur/things-eeg2 ↗](https://huggingface.co/datasets?search=things-eeg2) — Hugging Face · 86 files · 147,944,847,435 bytes · revision `0e2ad98` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0178) | | [graycatHCO3/BCICIV_2A_tensor ↗](https://huggingface.co/datasets?search=BCICIV_2A_tensor) — Hugging Face · 4 files · 873,061,971 bytes · revision `4a4178b` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches BCI Competition IV data sets in dataset name and file names ([terms page captured 19 October 2026 ↗](https://www.bbci.de/competition/iv/)).; Matches BNCI Horizon 2020 data sets catalogue in dataset name and file names ([terms page captured 19 October 2026 ↗](https://bnci-horizon-2020.eu/database/data-sets)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0179) | | [graycatHCO3/BCICIV_2B ↗](https://huggingface.co/datasets?search=BCICIV_2B) — Hugging Face · 3 files · 234,811,217 bytes · revision `621e8d4` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0180) | | [graycatHCO3/BI_BCI2000_MELR ↗](https://huggingface.co/datasets?search=BI_BCI2000_MELR) — Hugging Face · 6 files · 271,748,783 bytes · revision `2dc1294` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0181) | | [graycatHCO3/bi_bci2000_MELR_tensor ↗](https://huggingface.co/datasets?search=bi_bci2000_MELR_tensor) — Hugging Face · 4 files · 1,538,523,797 bytes · revision `103fa8c` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches physionet-eegmmidb-1.0.0 in file names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0182) | | [graycatHCO3/comp90051-eegmmidb-prepared-20261014 ↗](https://huggingface.co/datasets?search=comp90051-eegmmidb-prepared-20261014) — Hugging Face · 21 files · 6,164,180,502 bytes · revision `4e2b01c` | — | `odc-by` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0183) | | [graycatHCO3/eegmmidb_0to8_cross ↗](https://huggingface.co/datasets?search=eegmmidb_0to8_cross) — Hugging Face · 108 files · 5,640,370,929 bytes · revision `8f41187` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches physionet-eegmmidb-1.0.0 in dataset name and file names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0184) | | [graycatHCO3/eegmmidb_denoised ↗](https://huggingface.co/datasets?search=eegmmidb_denoised) — Hugging Face · 4 files · 10,923,930,602 bytes · revision `cbb7ac9` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0185) | | [graycatHCO3/eegmmidb_tensor ↗](https://huggingface.co/datasets?search=eegmmidb_tensor) — Hugging Face · 111 files · 21,391,137,123 bytes · revision `2d52322` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches physionet-eegmmidb-1.0.0 in dataset name and file names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `mit` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/eegmmidb/1.0.0/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0186) | | [gsbcilab/bci-dataset-v4 ↗](https://huggingface.co/datasets?search=bci-dataset-v4) — Hugging Face · 9,535 files · 346,810,129 bytes · revision `2be08b2` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0187) | | [gxyi/SEED_Processwd_EEG ↗](https://huggingface.co/datasets?search=SEED_Processwd_EEG) — Hugging Face · 2 files · 166,929,326 bytes · revision `a78590f` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0188) | | [gymprathap/Brain-Computer-Interaction-Eye-State-Classification ↗](https://huggingface.co/datasets?search=Brain-Computer-Interaction-Eye-State-Classification) — Hugging Face · 3 files · 1,694,181 bytes · revision `e83baf9` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches EEG Eye State (UCI Machine Learning Repository, id 264) in column names and example counts. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `apache-2.0` ([terms page captured 19 October 2026 ↗](https://archive.ics.uci.edu/dataset/264/eeg+eye+state)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0189) | | [hamedtu/eeg-motor-imagery-data ↗](https://huggingface.co/datasets?search=eeg-motor-imagery-data) — Hugging Face · 9 files · 326,712,120 bytes · revision `637e3e6` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0190) | | [heee565/SpanishBCBL ↗](https://huggingface.co/datasets?search=SpanishBCBL) — Hugging Face · revision `3bee883` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0191) | | [hezy18/EEG-SVRec ↗](https://huggingface.co/datasets?search=EEG-SVRec) — Hugging Face · 31 files · 63,782,652,030 bytes · revision `2597392` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0192) | | [iatosh/SleepEDF-V-Dataset ↗](https://huggingface.co/datasets?search=SleepEDF-V-Dataset) — Hugging Face · 6 files · 1,879,287,226 bytes · revision `9c32725` | — | `odc-by` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0193) | | [illusion-brain/primate-spike-20ms ↗](https://huggingface.co/datasets?search=primate-spike-20ms) — Hugging Face · 72 files · 2,403,825,941 bytes · revision `0e230bf` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0194) | | [introvoyz042/Egocentric-EMG-Force ↗](https://huggingface.co/datasets?search=Egocentric-EMG-Force) — Hugging Face · 14 files · 8,433,938,159 bytes · revision `3c2dcf5` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0195) | | [jalauer/Albrecht2019 ↗](https://huggingface.co/datasets?search=Albrecht2019) — Hugging Face · 377 files · 28,914,297,274 bytes · revision `53b7c87` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0196) | | [jalauer/Brown2020 ↗](https://huggingface.co/datasets?search=Brown2020) — Hugging Face · 143 files · 18,305,287,769 bytes · revision `b33cb61` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0197) | | [jalauer/Cavanagh2018a ↗](https://huggingface.co/datasets?search=Cavanagh2018a) — Hugging Face · 81 files · 12,337,929,879 bytes · revision `c00e673` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0198) | | [jalauer/Cavanagh2018b ↗](https://huggingface.co/datasets?search=Cavanagh2018b) — Hugging Face · 91 files · 2,088,813,253 bytes · revision `3859849` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0199) | | [jalauer/Cavanagh2019 ↗](https://huggingface.co/datasets?search=Cavanagh2019) — Hugging Face · 225 files · 48,634,141,402 bytes · revision `2b799ac` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0200) | | [jalauer/Gruendler2009 ↗](https://huggingface.co/datasets?search=Gruendler2009) — Hugging Face · 63 files · 7,385,925,457 bytes · revision `a2fa8aa` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0201) | | [jalauer/Singh2018 ↗](https://huggingface.co/datasets?search=Singh2018) — Hugging Face · 123 files · 100,235,121,847 bytes · revision `5e427d4` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0202) | | [jalauer/Singh2020 ↗](https://huggingface.co/datasets?search=Singh2020) — Hugging Face · 1,123 files · 4,068,828,564 bytes · revision `7a78668` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0203) | | [jalauer/Singh2021 ↗](https://huggingface.co/datasets?search=Singh2021) — Hugging Face · 391 files · 23,488,664,118 bytes · revision `85a6b39` | — | `pddl` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0204) | | [jamie33/July ↗](https://huggingface.co/datasets?search=July) — Hugging Face · 86,652 files · 365,826,927,648 bytes · revision `ed0109c` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0205) | | [jasontchan/2025-07-29-emg-pick-place-yellow-block ↗](https://huggingface.co/datasets?search=2025-07-29-emg-pick-place-yellow-block) — Hugging Face · 206 files · 1,384,705,978 bytes · revision `2dffb94` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0206) | | [jasontchan/emg-gestures-pick-place-correct-block-pp ↗](https://huggingface.co/datasets?search=emg-gestures-pick-place-correct-block-pp) — Hugging Face · 222 files · 1,522,037,405 bytes · revision `6b6bc7c` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0207) | | [jasontchan/emg-gestures-pick-place-yellow-block ↗](https://huggingface.co/datasets?search=emg-gestures-pick-place-yellow-block) — Hugging Face · 206 files · 1,397,309,297 bytes · revision `c7e064c` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0208) | | [jasontchan/emg-gestures-pick-place-yellow-block-new ↗](https://huggingface.co/datasets?search=emg-gestures-pick-place-yellow-block-new) — Hugging Face · 14 files · 35,472,181 bytes · revision `3008e4e` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0209) | | [jasontchan/emg-gestures-pick-place-yellow-block-pp ↗](https://huggingface.co/datasets?search=emg-gestures-pick-place-yellow-block-pp) — Hugging Face · 206 files · 1,532,534,419 bytes · revision `8558043` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0210) | | [jasontchan/emg-gestures-task3-edit ↗](https://huggingface.co/datasets?search=emg-gestures-task3-edit) — Hugging Face · 198 files · 1,983,911,643 bytes · revision `f99d0aa` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0216) | | [jasontchan/emg-pick-place-yellow-block ↗](https://huggingface.co/datasets?search=emg-pick-place-yellow-block) — Hugging Face · 159 files · 1,510,889,044 bytes · revision `b588b29` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0218) | | [jasontchan/emg-pick-place-yellow-block-pp ↗](https://huggingface.co/datasets?search=emg-pick-place-yellow-block-pp) — Hugging Face · 156 files · 1,368,226,723 bytes · revision `597ebb2` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0219) | | [jasontchan/emg-test ↗](https://huggingface.co/datasets?search=emg-test) — Hugging Face · 15 files · 143,786,647 bytes · revision `9a6855c` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0220) | | [jasontchan/pick-place-three-blocks-full-emg-corrected ↗](https://huggingface.co/datasets?search=pick-place-three-blocks-full-emg-corrected) — Hugging Face · 60 files · 1,954,052,600 bytes · revision `268079b` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0231) | | [jatshi/CAS_EEG ↗](https://huggingface.co/datasets?search=CAS_EEG) — Hugging Face · 16 files · 22,321,373,716 bytes · revision `8a18106` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0234) | | [jlgr/EMG-Dataset ↗](https://huggingface.co/datasets?search=EMG-Dataset) — Hugging Face · 19,767 files · 675,553 bytes · revision `8364765` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0235) | | [jomonjoseph98/real-eeg ↗](https://huggingface.co/datasets?search=real-eeg) — Hugging Face · 7 files · 531,293 bytes · revision `97dd8af` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0236) | | [jonathan-lys/reve-atlas ↗](https://huggingface.co/datasets?search=reve-atlas) — Hugging Face · 54 files · 3,182,997,399 bytes · revision `a1d6ca2` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0237) | | [justkha05/EEG ↗](https://huggingface.co/datasets?search=EEG) — Hugging Face · 397 files · 8,715,150,753 bytes · revision `d593938` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0238) | | [kgarg47/kahlus-raw-eeg-waveforms-efa0d90 ↗](https://huggingface.co/datasets?search=kahlus-raw-eeg-waveforms-efa0d90) — Hugging Face · 4,378 files · 79,888,820,959 bytes · revision `2e93a74` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0239) | | [kkchaulagain/LieWaves-EEG ↗](https://huggingface.co/datasets?search=LieWaves-EEG) — Hugging Face · 3 files · 6,991,046 bytes · revision `b5cfeab` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0240) | | [kristinbranson/neurodata-reuse-zhang2025 ↗](https://huggingface.co/datasets?search=neurodata-reuse-zhang2025) — Hugging Face · 1,750 files · 49,440,663,334 bytes · revision `b134ad2` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0241) | | [lee-jhwn/multimodal-speech-biosignals ↗](https://huggingface.co/datasets?search=multimodal-speech-biosignals) — Hugging Face · 369 files · 1,189,600,689 bytes · revision `2e36764` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0243) | | [lemon2004/DREAMS ↗](https://huggingface.co/datasets?search=DREAMS) — Hugging Face · 19 files · 4,259,338,561 bytes · revision `b131c6d` | — | `cc-by-nc-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0244) | | [liverman/EEG_DECODE ↗](https://huggingface.co/datasets?search=EEG_DECODE) — Hugging Face · 16 files · 21,374,064,363 bytes · revision `5a91bb9` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches THINGS-EEG2 (Gifford et al. 2022) in file hashes. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `apache-2.0` ([terms page captured 19 October 2026 ↗](https://osf.io/3jk45/)).; Matches THINGS object image database in file hashes ([terms page captured 19 October 2026 ↗](https://osf.io/download/ahg2k/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0245) | | [liyongsea/THINGS_EEG_Test ↗](https://huggingface.co/datasets?search=THINGS_EEG_Test) — Hugging Face · 13 files · 2,444,622,232 bytes · revision `6f69761` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0246) | | [luigi-s/EEG_Image_CVPR_ALL_subj ↗](https://huggingface.co/datasets?search=EEG_Image_CVPR_ALL_subj) — Hugging Face · 12 files · 1,705,330,595 bytes · revision `b8e5c68` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0247) | | [luigi-s/EEG_Image_TVIZ_ALL_subj ↗](https://huggingface.co/datasets?search=EEG_Image_TVIZ_ALL_subj) — Hugging Face · 37 files · 13,164,881,638 bytes · revision `1eba80b` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0248) | | [madteam/thinking-bciciv2a ↗](https://huggingface.co/datasets?search=thinking-bciciv2a) — Hugging Face · 4 files · 214,854,413 bytes · revision `1cd0dea` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0249) | | [malansi/EMG-Gesture-Dataset ↗](https://huggingface.co/datasets?search=EMG-Gesture-Dataset) — Hugging Face · 10 files · 12,080,547 bytes · revision `27dc1e5` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0250) | | [marical/eeg-eye-state ↗](https://huggingface.co/datasets?search=eeg-eye-state) — Hugging Face · 2 files · 1,762,308 bytes · revision `171784a` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0251) | | [matvelen6369/eeg_data ↗](https://huggingface.co/datasets?search=eeg_data) — Hugging Face · 15 files · 4,934,972,199 bytes · revision `51edd5b` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0252) | | [matvelen6369/eeg_data_disoder ↗](https://huggingface.co/datasets?search=eeg_data_disoder) — Hugging Face · 11 files · 2,891,512,799 bytes · revision `240c388` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0253) | | [mod1155/CineBrain ↗](https://huggingface.co/datasets?search=CineBrain) — Hugging Face · 19 files · 140,971,724,389 bytes · revision `31bbf1f` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0254) | | [monster-monash/CrowdSourced ↗](https://huggingface.co/datasets?search=CrowdSourced) — Hugging Face · 16 files · 918,541,922 bytes · revision `18ab93f` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0255) | | [monster-monash/DREAMERA ↗](https://huggingface.co/datasets?search=DREAMERA) — Hugging Face · 12 files · 2,444,489,648 bytes · revision `7a55408` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Its card names DREAMER as its source.; Matches DREAMER in the source its card names. That source is available on request; this upload’s licence tag is `other` ([terms page captured 19 October 2026 ↗](https://zenodo.org/records/546113)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0256) | | [monster-monash/DREAMERV ↗](https://huggingface.co/datasets?search=DREAMERV) — Hugging Face · 11 files · 2,443,127,553 bytes · revision `2bfd78c` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Its card names DREAMER as its source.; Matches DREAMER in the source its card names. That source is available on request; this upload’s licence tag is `other` ([terms page captured 19 October 2026 ↗](https://zenodo.org/records/546113)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0257) | | [musicji83/amigos-eeg-dataset ↗](https://huggingface.co/datasets?search=amigos-eeg-dataset) — Hugging Face · 41 files · 3,877,621,403 bytes · revision `f3234f6` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Matches AMIGOS in byte counts, file count, and file names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://www.eecs.qmul.ac.uk/mmv/datasets/amigos/download.html)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0259) | | [mysteriousauthor/spikeprophecy-steinmetz ↗](https://huggingface.co/datasets?search=spikeprophecy-steinmetz) — Hugging Face · 43 files · 1,761,488,739 bytes · revision `04b11ea` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0260) | | [nagyadam97/EEG_records_raw_schizophrenia_bipolar ↗](https://huggingface.co/datasets?search=EEG_records_raw_schizophrenia_bipolar) — Hugging Face · 88 files · 26,985,895,091 bytes · revision `17917b6` | — | `cc-by-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0261) | | [naplabdataset/mov-aad ↗](https://huggingface.co/datasets?search=mov-aad) — Hugging Face · 165 files · 319,780,688,436 bytes · revision `45e7dde` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0262) | | [nariow/EEG-Electrode-Shift-Dataset ↗](https://huggingface.co/datasets?search=EEG-Electrode-Shift-Dataset) — Hugging Face · 42 files · 29,035,639,837 bytes · revision `c7cf8f7` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0263) | | [neukym/kymata-soto ↗](https://huggingface.co/datasets?search=kymata-soto) — Hugging Face · 715 files · 174,637,019,414 bytes · revision `dd4cc91` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0264) | | [neurodeskorg/eeg-frequency-tagging-sample ↗](https://huggingface.co/datasets?search=eeg-frequency-tagging-sample) — Hugging Face · 4 files · 157,379,440 bytes · revision `b04151e` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0265) | | [neurofusion/eeg-restingstate ↗](https://huggingface.co/datasets?search=eeg-restingstate) — Hugging Face · 155 files · 694,723,700 bytes · revision `fe43ac7` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0266) | | [neuroinfinite/fnirs-finekyber ↗](https://huggingface.co/datasets?search=fnirs-finekyber) — Hugging Face · 31 files · 1,410,808,755 bytes · revision `9405799` | — | `cc-by-nc-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches Guevara et al. 2024 fNIRS motor-cortex dataset (Parkinson's disease) in the source its card names. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `cc-by-nc-sa-4.0` ([terms page captured 19 October 2026 ↗](https://zenodo.org/records/7966830)).; Matches Multiple-session NIRS recordings from stroke patients controlling a BCI (Isaev et al. 2024) in the source its card names. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `cc-by-nc-sa-4.0` ([terms page captured 19 October 2026 ↗](https://www.nitrc.org/doi/landing_page.php?table=groups&id=1650&doi=10.25790/bml0cm.167)).; Matches ASTaUND fNIRS dataset (Kamat et al. 2023) in the source its card names. That source’s terms page gives `cc0-1.0`; this upload’s licence tag is `cc-by-nc-sa-4.0` ([terms page captured 19 October 2026 ↗](https://doi.org/10.6084/m9.figshare.23643138.v1)).; Matches fNIRS Finger Tapping Data in BIDS Format (Luke et al.) in the source its card names ([terms page captured 19 October 2026 ↗](https://zenodo.org/records/5529797)).; Matches fNIRS resting state with motion artifacts (Yucel et al. 2014) in the source its card names. That source’s terms page gives `bsd`; this upload’s licence tag is `cc-by-nc-sa-4.0` ([terms page captured 19 October 2026 ↗](https://www.nitrc.org/projects/yucel14motion/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0267) | | [omarfirdaus101/emg-cyton-digits ↗](https://huggingface.co/datasets?search=emg-cyton-digits) — Hugging Face · 6 files · 70,637,742 bytes · revision `af097c4` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0268) | | [opsecsystems/D1D2_Fundamental ↗](https://huggingface.co/datasets?search=D1D2_Fundamental) — Hugging Face · 974 files · 6,832,718,344 bytes · revision `4e4a0d1` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0269) | | [opsecsystems/eegdream ↗](https://huggingface.co/datasets?search=eegdream) — Hugging Face · 3 files · 12,856,975 bytes · revision `d1d5978` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0270) | | [perottofederico/EEGCVPR40-captions ↗](https://huggingface.co/datasets?search=EEGCVPR40-captions) — Hugging Face · 12 files · 1,746,881,045 bytes · revision `6e26220` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0271) | | [perottofederico/EEGCVPR40_GoodCaptions ↗](https://huggingface.co/datasets?search=EEGCVPR40_GoodCaptions) — Hugging Face · 50 files · 21,156,860,899 bytes · revision `c13cf7c` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0272) | | [perottofederico/EEGCVPR40_good_captions ↗](https://huggingface.co/datasets?search=EEGCVPR40_good_captions) — Hugging Face · 12 files · 1,704,243,765 bytes · revision `8ffbf56` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0273) | | [perottofederico/things-eeg ↗](https://huggingface.co/datasets?search=things-eeg) — Hugging Face · 162 files · 54,194,614,603 bytes · revision `2850daa` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0274) | | [phuongbtmp/EEG2Speech ↗](https://huggingface.co/datasets?search=EEG2Speech) — Hugging Face · 4,807 files · 70,761,809 bytes · revision `72a8d33` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0275) | | [prastab/eeg-processed-train-dataset ↗](https://huggingface.co/datasets?search=eeg-processed-train-dataset) — Hugging Face · 78 files · 14,261,009,488 bytes · revision `89cb03d` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0276) | | [praxilon76/Sample_data ↗](https://huggingface.co/datasets?search=Sample_data) — Hugging Face · 16,645 files · 8,581,997,442 bytes · revision `7d0b139` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: access is granted automatically once a user accepts the conditions. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0277) | | [pyinglie/EEG-To-Text_Datasets ↗](https://huggingface.co/datasets?search=EEG-To-Text_Datasets) — Hugging Face · 78 files · 69,447,746,815 bytes · revision `97e14eb` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0278) | | [pyinglie/EEG-To-Text_Preprocessed ↗](https://huggingface.co/datasets?search=EEG-To-Text_Preprocessed) — Hugging Face · 9 files · 5,111,299,297 bytes · revision `af5f576` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0279) | | [raei/Brunner-2008-BCICompetition2008GrazDataSetA ↗](https://huggingface.co/datasets?search=Brunner-2008-BCICompetition2008GrazDataSetA) — Hugging Face · 22 files · 780,098,116 bytes · revision `a51c59d` | — | `cc-by-nd-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0280) | | [raei/Jeong-2022-InternationalBCICompetition2020Review-track3 ↗](https://huggingface.co/datasets?search=Jeong-2022-InternationalBCICompetition2020Review-track3) — Hugging Face · 52 files · 2,326,563,932 bytes · revision `60718f1` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names BCIComp2020 as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0281) | | [raei/Khanday-2026-EEGFoundationModelsSpeechDecodingBenchmark ↗](https://huggingface.co/datasets?search=Khanday-2026-EEGFoundationModelsSpeechDecodingBenchmark) — Hugging Face · 536 files · 22,025,725,694 bytes · revision `b87226f` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0282) | | [raei/Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses ↗](https://huggingface.co/datasets?search=Kunz-2025-InnerSpeechMotorCortexImplicationsSpeechNeuroprostheses) — Hugging Face · 84 files · 11,500,388,137 bytes · revision `3f4006d` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0283) | | [raei/Nieto-2022-ThinkingOutLoudOpenAccessEEGBasedBCIDatasetInnerSpeech ↗](https://huggingface.co/datasets?search=Nieto-2022-ThinkingOutLoudOpenAccessEEGBasedBCIDatasetInnerSpeech) — Hugging Face · 185 files · 26,828,601,779 bytes · revision `e0a07f4` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | Its card names Nieto2022 as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0284) | | [raei/Schalk-2009-EEGMotorMovementImagery ↗](https://huggingface.co/datasets?search=Schalk-2009-EEGMotorMovementImagery) — Hugging Face · 3,061 files · 3,614,974,014 bytes · revision `d9c07d0` | — | `odc-by` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0285) | | [rishab-iyer1/mindeye_ieeg ↗](https://huggingface.co/datasets?search=mindeye_ieeg) — Hugging Face · 16,917 files · 55,331,509,685 bytes · revision `63315ee` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Matches Natural Scenes Dataset (NSD) in file names. That source requires registration; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://cvnlab.slite.page/p/IB6BSeW_7o/Terms-and-Conditions)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0286) | | [riteshhf/tsfp-repro-adftd-processed ↗](https://huggingface.co/datasets?search=tsfp-repro-adftd-processed) — Hugging Face · 12 files · 1,358,220,335 bytes · revision `74286a5` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0287) | | [rverma0631/eeg40 ↗](https://huggingface.co/datasets?search=eeg40) — Hugging Face · 2 files · 18,354,457,929 bytes · revision `240a2d0` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0288) | | [sajjad5221/eeg2text-emmt-dataset ↗](https://huggingface.co/datasets?search=eeg2text-emmt-dataset) — Hugging Face · 5 files · 19,876,530 bytes · revision `4d6a44b` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0289) | | [sajjad5221/eeg2text-emmt-dataset-unique-sentences ↗](https://huggingface.co/datasets?search=eeg2text-emmt-dataset-unique-sentences) — Hugging Face · 5 files · 36,048,867 bytes · revision `09caa21` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0290) | | [sajjad5221/emmt-eeg-raw-phases ↗](https://huggingface.co/datasets?search=emmt-eeg-raw-phases) — Hugging Face · 7 files · 90,833,978 bytes · revision `9806cb2` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0291) | | [sarveshrbabu/eeg-seizure-data ↗](https://huggingface.co/datasets?search=eeg-seizure-data) — Hugging Face · 221 files · 13,334,989,115 bytes · revision `fddd8dc` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0292) | | [skywalker-p/EEG_2d ↗](https://huggingface.co/datasets?search=EEG_2d) — Hugging Face · 121 files · 15,360,017,821 bytes · revision `3a71cc9` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0293) | | [skywalker-p/PAIR ↗](https://huggingface.co/datasets?search=PAIR) — Hugging Face · 2,171 files · 8,114,705,448 bytes · revision `73d3330` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0294) | | [smam/chb-mit-seizure-windows-2to1 ↗](https://huggingface.co/datasets?search=chb-mit-seizure-windows-2to1) — Hugging Face · 8 files · 3,046,430,860 bytes · revision `964002e` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches CHB-MIT Scalp EEG Database (PhysioNet chbmit 1.0.0) in the source its card names. That source’s terms page gives `odc-by-1.0`; this upload’s licence tag is `cc-by-4.0` ([terms page captured 19 October 2026 ↗](https://physionet.org/content/chbmit/1.0.0/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0295) | | [smam/tuh-eeg ↗](https://huggingface.co/datasets?search=tuh-eeg) — Hugging Face · 11,128 files · 2,175,495,133,396 bytes · revision `2d50a5e` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | No licence tag on its Hugging Face listing.; Gated on Hugging Face: the uploader approves each request.; Matches TUH EEG Corpus in folder names. That source is released under a signed agreement; this upload has no licence tag ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0296) | | [srimeenakshiks/eeg-biometric-data ↗](https://huggingface.co/datasets?search=eeg-biometric-data) — Hugging Face · 176 files · 425,887,560 bytes · revision `297c302` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0297) | | [strawcoder33/bci-mamba-data ↗](https://huggingface.co/datasets?search=bci-mamba-data) — Hugging Face · 142 files · 25,044,062,776 bytes · revision `f135d98` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0298) | | [tankalapavankalyan/exp01-eeg-to-text-sentences ↗](https://huggingface.co/datasets?search=exp01-eeg-to-text-sentences) — Hugging Face · 127 files · 71,606,968,889 bytes · revision `957f0f2` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0299) | | [tobisns/multisource-eeg-motorimagery ↗](https://huggingface.co/datasets?search=multisource-eeg-motorimagery) — Hugging Face · 7 files · 4,881,719,195 bytes · revision `57a6412` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0300) | | [tqxg2022/EEG3D ↗](https://huggingface.co/datasets?search=EEG3D) — Hugging Face · 29 files · 76,579,598,133 bytes · revision `3512601` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0301) | | tttgxias/Omni-iEEG — Hugging Face · 4,270 files · 159,246,369,342 bytes · revision `c2fb59e` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0302) | | [ut-vision/EgoBrain-Mini ↗](https://huggingface.co/datasets?search=EgoBrain-Mini) — Hugging Face · 17,520 files · 44,568,261,067 bytes · revision `dfd2906` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0303) | | [velvet-pine-22/PULSE ↗](https://huggingface.co/datasets?search=PULSE) — Hugging Face · 3,623 files · 78,365,484,728 bytes · revision `5c87906` | — | `cc-by-nc-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0304) | | [wang124/L-Mind ↗](https://huggingface.co/datasets?search=L-Mind) — Hugging Face · 12,012 files · 2,422,510,990 bytes · revision `6557307` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0305) | | [wbxlala/sleep_edf_3he1 ↗](https://huggingface.co/datasets?search=sleep_edf_3he1) — Hugging Face · 4 files · 102,253,079 bytes · revision `999708c` | — | `odc-by` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0308) | | [wbxlala/sleep_edf_str ↗](https://huggingface.co/datasets?search=sleep_edf_str) — Hugging Face · 4 files · 309,018,401 bytes · revision `1c77f75` | — | `odc-by` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0309) | | [wd41/isruc-sleep ↗](https://huggingface.co/datasets?search=isruc-sleep) — Hugging Face · 1,392 files · 13,105,355,460 bytes · revision `85bae4c` | — | `unknown` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0310) | | [whhhy123/EEGAgentBenchmark ↗](https://huggingface.co/datasets?search=EEGAgentBenchmark) — Hugging Face · 1,041 files · 24,244,825,183 bytes · revision `7df779a` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0311) | | [windwerfer/neurofeed-eeg-windows ↗](https://huggingface.co/datasets?search=neurofeed-eeg-windows) — Hugging Face · 1,700 files · 2,738,924,606 bytes · revision `43e5bc9` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0312) | | [woldier/eeg_denoise_dataset ↗](https://huggingface.co/datasets?search=eeg_denoise_dataset) — Hugging Face · 6 files · 1,871,560,176 bytes · revision `1835df9` | — | `apache-2.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Matches EEGdenoiseNet (NCClab, Southern University of Science and Technology) in the source its card names. That source’s terms page gives `mit`; this upload’s licence tag is `apache-2.0` ([terms page captured 19 October 2026 ↗](https://github.com/ncclabsustech/EEGdenoiseNet/blob/8d290661146c7189c98cc04812d37371d4b9426c/LICENSE)).; Matches Semi-simulated EEG/EOG dataset for EOG artifact rejection (Klados and Bamidis 2016) in the source its card names. That source’s terms page gives `cc-by-4.0`; this upload’s licence tag is `apache-2.0` ([terms page captured 19 October 2026 ↗](https://data.mendeley.com/datasets/wb6yvr725d/1)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0313) | | [xkoo115/UnifiedNeuroGen-Demo-Dataset ↗](https://huggingface.co/datasets?search=UnifiedNeuroGen-Demo-Dataset) — Hugging Face · 291 files · 7,108,055,842 bytes · revision `aa85843` | — | `mit` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0314) | | [xw1216/openneuro-ds005178-eesm23-v1-0-0-transfer ↗](https://huggingface.co/datasets?search=openneuro-ds005178-eesm23-v1-0-0-transfer) — Hugging Face · 10 files · 41,598,867,606 bytes · revision `04474ed` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request.; Its card names EESM23 · Ear-EEG Sleep Monitoring 2023 as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0315) | | [xw1216/openneuro-ds005185-eesm19-raw-v1-0-2 ↗](https://huggingface.co/datasets?search=openneuro-ds005185-eesm19-raw-v1-0-2) — Hugging Face · 3,556 files · 841,008,537,323 bytes · revision `1dbffee` | — | `cc0-1.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request.; Its card names EESM19 scalp subset as its source. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0316) | | [xw1216/openneuro-hbn-eeg-releases-v1-0-1 ↗](https://huggingface.co/datasets?search=openneuro-hbn-eeg-releases-v1-0-1) — Hugging Face · 15 files · 1,748,294,708,998 bytes · revision `8b6c69a` | — | `cc-by-sa-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | Gated on Hugging Face: the uploader approves each request. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0317) | | [xw1216/tuh-eeg-cloud-transfer ↗](https://huggingface.co/datasets?search=tuh-eeg-cloud-transfer) — Hugging Face · 286 files · 1,756,624,868,397 bytes · revision `32fdab0` | — | `other` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the re-upload rule | Gated on Hugging Face: the uploader approves each request.; Its card names TUH EEG Corpus as its source.; Matches TUH EEG Corpus in the source its card names. That source is released under a signed agreement; this upload’s licence tag is `other` ([terms page captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/)). [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0318) | | [yefllower/EEGFaceSem ↗](https://huggingface.co/datasets?search=EEGFaceSem) — Hugging Face · 187 files · 26,200,306,457 bytes · revision `49930c9` | — | `cc-by-4.0` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0319) | | [yushiran/fnirs_fmri_stroop ↗](https://huggingface.co/datasets?search=fnirs_fmri_stroop) — Hugging Face · 2 files · 803,958,814 bytes · revision `ee72b66` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0320) | | [yushiran/fnirs_stroop ↗](https://huggingface.co/datasets?search=fnirs_stroop) — Hugging Face · 2 files · 291,442 bytes · revision `e0bc801` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0321) | | [zl3030/reliant-bci-nlb-public-raw ↗](https://huggingface.co/datasets?search=reliant-bci-nlb-public-raw) — Hugging Face · 2 files · 2,209,607,742 bytes · revision `e078a5d` | — | `none` — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | No licence tag on its Hugging Face listing. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0322) | | [SanderGi/lerobot-eeg-test* ↗](https://huggingface.co/datasets?search=lerobot-eeg-test) — Hugging Face · 1,106,524,422 bytes in all | — | `apache-2.0` (22) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 22 repositories by SanderGi in the series “lerobot-eeg-test”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0323) | | [dbzfan2012/lerobot-eeg-test* ↗](https://huggingface.co/datasets?search=lerobot-eeg-test) — Hugging Face · 97,479,718 bytes in all | — | `none` (5), `apache-2.0` (1) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 6 repositories by dbzfan2012 in the series “lerobot-eeg-test”, listed as one row.; 5 of 6 members carry no licence tag. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0324) | | [perottofederico/things-eeg-captions* ↗](https://huggingface.co/datasets?search=things-eeg-captions) — Hugging Face · 345,309,729,711 bytes in all | — | `none` (13) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 13 repositories by perottofederico in the series “things-eeg-captions”, listed as one row.; 13 of 13 members carry no licence tag. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0325) | | [tankalapavankalyan/ds007808* ↗](https://huggingface.co/datasets?search=ds007808) — Hugging Face · 330,175,482,488 bytes in all | — | `cc0-1.0` (1) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 1 repository by tankalapavankalyan in the series “ds007808”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0326) | | [Kkuntal990/bnci* ↗](https://huggingface.co/datasets?search=bnci) — Hugging Face · 22,800,659 bytes in all | — | `unknown` (4) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 4 repositories by Kkuntal990 in the series “bnci”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0327) | | [Kkuntal990/test* ↗](https://huggingface.co/datasets?search=test) — Hugging Face · 22,800,659 bytes in all | — | `unknown` (4) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 4 repositories by Kkuntal990 in the series “test”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0328) | | [SanderGi/eval_lerobot-eeg-test* ↗](https://huggingface.co/datasets?search=eval_lerobot-eeg-test) — Hugging Face · 6,252,710 bytes in all | — | `apache-2.0` (1), `none` (1) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 2 repositories by SanderGi in the series “eval_lerobot-eeg-test”, listed as one row.; 1 of 2 members carries no licence tag. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0329) | | [atgarcia/EMGSoundTrain* ↗](https://huggingface.co/datasets?search=EMGSoundTrain) — Hugging Face · 3,200,955,982 bytes in all | — | `none` (4) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 4 repositories by atgarcia in the series “EMGSoundTrain”, listed as one row.; 4 of 4 members carry no licence tag. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0330) | | [braindecode/example_dataset* ↗](https://huggingface.co/datasets?search=example_dataset) — Hugging Face · 21,800,049 bytes in all | — | `unknown` (3) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 3 repositories by braindecode in the series “example_dataset”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0331) | | [jasontchan/emg-gestures-task* ↗](https://huggingface.co/datasets?search=emg-gestures-task) — Hugging Face · 10,545,603,049 bytes in all | — | `apache-2.0` (6) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 6 repositories by jasontchan in the series “emg-gestures-task”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0332) | | [jasontchan/eval_emg-task1-bigger* ↗](https://huggingface.co/datasets?search=eval_emg-task1-bigger) — Hugging Face · 1,661,249,331 bytes in all | — | `apache-2.0` (5) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 5 repositories by jasontchan in the series “eval_emg-task1-bigger”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0333) | | [jasontchan/eval_emg-task2-big* ↗](https://huggingface.co/datasets?search=eval_emg-task2-big) — Hugging Face · 97,289,643 bytes in all | — | `apache-2.0` (1) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 1 repository by jasontchan in the series “eval_emg-task2-big”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0334) | | [jasontchan/eval_emg-task3-edit* ↗](https://huggingface.co/datasets?search=eval_emg-task3-edit) — Hugging Face · 739,729,794 bytes in all | — | `apache-2.0` (4) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 4 repositories by jasontchan in the series “eval_emg-task3-edit”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0335) | | [jasontchan/testing* ↗](https://huggingface.co/datasets?search=testing) — Hugging Face · 1,777,410,661 bytes in all | — | `apache-2.0` (2) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 2 repositories by jasontchan in the series “testing”, listed as one row. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0336) | | [wbxlala/eegimage* ↗](https://huggingface.co/datasets?search=eegimage) — Hugging Face · 811,602,675 bytes in all | — | `none` (2) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 2 repositories by wbxlala in the series “eegimage”, listed as one row.; 2 of 2 members carry no licence tag. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-hf-0337) | | **OpenNeuro via EEGDash** · 1 row — EEGDash’s mirrors of OpenNeuro and NEMAR records, as one row. Which mirror holds which record is in the JSON file; this site’s rows and MOABB rows that have a mirror list it among their names. | | | | | | OpenNeuro and NEMAR via EEGDash — Hugging Face — Also: EEGDash | — | `cc0-1.0` (553), `cc-by-4.0` (91), `other` (46), `cc-by-sa-4.0` (37), `cc-by-nc-4.0` (6), `cc-by-nc-sa-4.0` (4), `cc-by-nc-nd-4.0` (1), `cc-by-nd-4.0` (1), `none` (1), `odc-by` (1) — Hugging Face licence tag | Step 1 of 5 · Listed — Status: Listed | 737 repositories by EEGDash in the series “OpenNeuro and NEMAR mirrors”, listed as one row.; 1 of 737 members carries no licence tag. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-openneuro-mirrors) | | **This site’s datasets** · 52 rows — Datasets this site has scored, held or reviewed, furthest step first. A row with results links its page. | | | | | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — Also: BETA database, The BETA database, BETA: A Large Benchmark Database Toward SSVEP-BCI Application, Liu et al. (2020) · MOABB: Liu2020BETA · DOI: 10.3389/fnins.2020.00627 · Hugging Face: Bingchuan/BETA | 40-target SSVEP — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 5 of 5 · In the matrix — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/beta/) · [Named in a pretraining corpus? →](https://bci.report/protocols/beta-8ch/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-beta) | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Also: Motor Imagery vs Rest - Low-Cost EEG System, A motor imagery vs. rest dataset with low-cost consumer grade EEG hardware, Peterson et al. (2022) · MOABB: Peterson2020 · DOI: 10.1016/j.heliyon.2020.e03425 · EEGDash: EEGDash/ds003810 | Motor imagery / rest — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 5 of 5 · In the matrix — share above chance: not computed — Status: Published | LaBraM at Scale Atom-L and LaBraM at Scale Molecule-L list it in their pretraining corpus; which release of it they used is not established. [On this site →](https://bci.report/datasets/ds003810/) · [Named in a pretraining corpus? →](https://bci.report/protocols/mi-rest/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ds003810) | | [ds005342 ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) — Also: EEG data offline and online during motor imagery for standing and sitting, Triana-Guzman et al. (2022) · MOABB: TrianaGuzman2024 · DOI: 10.3389/fninf.2022.961089 · EEGDash: EEGDash/ds005342 | Cue-gated idle / command — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 5 of 5 · In the matrix — Status: Published | LaBraM at Scale Atom-L and LaBraM at Scale Molecule-L list it in their pretraining corpus; which release of it they used is not established. [On this site →](https://bci.report/datasets/ds005342/) · [Named in a pretraining corpus? →](https://bci.report/protocols/idle/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ds005342) | | [ds006593 ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) — Also: cBCI Matrix Multimodal Dataset, Multimodal Sensor Fusion for EEG-Based BCI Typing Systems, on006593 · EEGDash: EEGDash/ds006593 | P300 target ERP — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 5 of 5 · In the matrix — share above chance: not computed — Status: Published | LaBraM at Scale Atom-L and LaBraM at Scale Molecule-L list it in their pretraining corpus; which release of it they used is not established. [On this site →](https://bci.report/datasets/ds006593/) · [Named in a pretraining corpus? →](https://bci.report/protocols/p300-target/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ds006593) | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Also: EEG During Mental Arithmetic Tasks, Electroencephalograms during Mental Arithmetic Task Performance, Zyma et al. (2019) | Mental arithmetic / rest — as this site’s page states it | Open Data Commons Attribution License 1.0 — this site’s rights record | Step 5 of 5 · In the matrix — share above chance: not computed — Status: Published | LaBraM at Scale Atom-L and LaBraM at Scale Molecule-L list it in their pretraining corpus; which release of it they used is not established. [On this site →](https://bci.report/datasets/eegmat/) · [Named in a pretraining corpus? →](https://bci.report/protocols/arithmetic-rest/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-eegmat) | | [EESM19 scalp subset ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Also: EESM19, Ear-EEG Sleep Monitoring 2019, ds005185, Mikkelsen et al. (2019) · EEGDash: EEGDash/ds005185 | Five-stage sleep — as this site’s page states it | CC0-1.0 declared by upstream and mirror — this site’s rights record | Step 5 of 5 · In the matrix — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/eesm19/) · [Named in a pretraining corpus? →](https://bci.report/protocols/sleep-scalp/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-eesm19) | | [TMNRED / ds005383 ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) — Also: TMNRED, ds005383, TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments, Bai et al. (2025) · EEGDash: EEGDash/ds005383 | Semantic target ERP — as this site’s page states it | OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 — this site’s rights record | Step 5 of 5 · In the matrix — share above chance: not computed — Status: Published | LaBraM at Scale Atom-L and LaBraM at Scale Molecule-L list it in their pretraining corpus; which release of it they used is not established. [On this site →](https://bci.report/datasets/tmnred/) · [Named in a pretraining corpus? →](https://bci.report/protocols/semantic-target/#pretraining-exposure) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-tmnred) | | [EEG Alpha Waves dataset ↗](https://doi.org/10.5281/zenodo.2605110) — Also: ALPHA.EEG.2017-GIPSA, Zenodo 2605110, hal-02086581 · MOABB: Rodrigues2017 · DOI: 10.5281/zenodo.2348891 | Eyes open / eyes closed — as this site’s page states it | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 3: consent or ethics statement — share above chance: not computed — Status: Published | The scores here use Zenodo record 2605110, which holds 20 recordings; 19 are scored. An earlier inventory named the source as Zenodo 2348892 / NEMAR nm000221. [On this site →](https://bci.report/datasets/alpha-waves/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-alpha-waves) | | [ASEEG · Atieh Schizophrenia EEG ↗](https://doi.org/10.5281/zenodo.18029536) — DOI: 10.5281/zenodo.18029536 | — | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/topics/clinical-groups/#aseeg) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-aseeg) | | [Bak et al. fNIRS finger- and foot-tapping dataset ↗](https://doi.org/10.6084/m9.figshare.9783755.v2) — Also: Open Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, figshare 9783755, fNIRS dataset for ternary classification, Bak et al. (2019) | Human fNIRS: right-hand, left-hand and foot tapping, a person left out at a time — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/bak-fnirs/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bak-fnirs) | | [BINGO cross-linguistic inner-speech dataset ↗](https://bingo-project.gr/datasets/) — Also: BINGO (Brain Imagined-Speech Communication), BINGO cross-linguistic inner speech elicitation dataset, BINGO Dataset 2, D3.1 – The BINGO Benchmarking Framework | Inner speech of 15 matched nouns in Greek and in English, across people and languages — as this site’s page states it | No licence stated — this site’s rights record | Step 4 of 5 · Run — gap at step 3: consent or ethics statement — gap at step 3: pinned version — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/bingo-imagined-speech/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bingo-imagined-speech) | | [BNCI2015-001 · next-day adaptation ↗](https://bnci-horizon-2020.eu/database/data-sets) — MOABB: BNCI2015_001 · DOI: 10.1109/tnsre.2012.2189584 | — | CC BY-NC-ND 4.0 (as displayed by the official catalogue) — this site’s rights record | Step 4 of 5 · Run — gap at step 3: rights review — gap at step 3: consent or ethics statement — gap at step 3: pinned version — gap at step 4: chance level — share above chance: not computed — Status: Score held | [The case →](https://bci.report/topics/model-adaptation/#next-day) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bnci2015-001) | | [BOAS · Bitbrain Open Access Sleep dataset ↗](https://openneuro.org/datasets/ds005555/versions/1.1.3) — Also: BOAS, Bitbrain Open Access Sleep dataset, The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, ds005555 · EEGDash: EEGDash/ds005555 | Overnight sleep, six polysomnography EEG channels, human-consensus stages — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/boas/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-boas) | | [Damm finger motor imagery · random and sequential order (OpenNeuro ds008446) ↗](https://doi.org/10.18112/openneuro.ds008446.v1.0.1) — MOABB: Damm2026 · DOI: 10.18112/openneuro.ds008446.v1.0.1 | — | CC0 — this site’s rights record | Step 4 of 5 · Run — gap at step 3: rights review — gap at step 3: consent or ethics statement — gap at step 4: chance level — Status: Score held | Held: with one person, every score is that person’s. [The case →](https://bci.report/topics/data-quality/#damm) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-damm-finger-imagery) | | [Dreem Open Datasets (DOD-H and DOD-O) ↗](https://zenodo.org/records/15900394) — Also: Dreem Open Datasets, DOD-H, DOD-O, DOD, Zenodo 15900394, Guillot et al. (2020) | Five-stage sleep staging, healthy sleepers and people with obstructive sleep apnoea, kept apart — as this site’s page states it | MIT — this site’s rights record | Step 4 of 5 · Run — gap at step 3: consent or ethics statement — gap at step 4: chance level — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/dreem-dod/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-dreem-dod) | | [OpenNeuro ds004584 · Parkinson's disease, rest eyes open ↗](https://doi.org/10.18112/openneuro.ds004584.v1.0.0) — Also: OpenNeuro ds004584, Rest eyes open, NEMAR on004584 · EEGDash: EEGDash/ds004584 | Resting-state EEG, Parkinson's disease and controls — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/ds004584/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ds004584) | | [EEG-fNIRS-based Handwriting Trajectory Dataset ↗](https://huggingface.co/datasets?search=EEG-fNIRS-based-Handwriting-Trajectory-Dataset) — Also: ACM MM 2026 Multimodal Brain-Computer Interface Grand Challenge · Hugging Face: lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset | Four released coded labels in an imagined-handwriting competition dataset, EEG only: a data-quality and methods case study — as this site’s page states it | CC BY-NC-SA 4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 3: derivation — gap at step 3: original terms — gap at step 3: consent or ethics statement — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/eeg-fnirs-handwriting/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-eeg-fnirs-handwriting) | | [EESM23 · Ear-EEG Sleep Monitoring 2023 ↗](https://doi.org/10.18112/openneuro.ds005178.v1.0.0) — Also: EESM23, Ear-EEG Sleep Monitoring 2023, OpenNeuro ds005178 · EEGDash: EEGDash/ds005178 | Five-stage sleep, in-ear and scalp — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/eesm23/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-eesm23) | | [eldBETA · eldercare-oriented SSVEP benchmark database ↗](https://doi.org/10.6084/m9.figshare.18032669.v3) — MOABB: Liu2022EldBETA · DOI: 10.1038/s41597-022-01372-9 | — | CC BY 4.0 — the deposit’s record | Step 4 of 5 · Run — Status: Published | SingLEM lists it in its pretraining corpus. [The case →](https://bci.report/topics/data-quality/#eldbeta) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-eldbeta) | | [Forenzo & He continuous-tracking EEG-BCI dataset ↗](https://doi.org/10.1184/R1/25360300.v1) — Also: Forenzo & He (2024), KiltHub 25360300, EEG-BCI Dataset for Continuous Tracking using Deep Learning-based Decoding for Non-invasive Brain-Computer Interface, Forenzo et al. (2024), continuous pursuit BCI dataset | Continuous cursor tracking with a noninvasive BCI, across sessions — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — Status: Published | [On this site →](https://bci.report/datasets/forenzo-continuous-tracking/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-forenzo-continuous-tracking) | | [Multimodal gait dataset · treadmill walking ↗](https://doi.org/10.13026/r0ea-7161) — Also: A multimodal gait dataset of brain activity, muscle activity, kinematics and ground forces in young adults, PhysioNet multimodal-gait-dataset, doi:10.13026/r0ea-7161 | Treadmill walking at three speeds — as this site’s page states it | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/gait-eeg/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-gait-eeg) | | [IESSEEG · infantile epileptic spasms syndrome EEG (Capur/IESSEEG) ↗](https://huggingface.co/datasets?search=IESSEEG) — Hugging Face: Capur/IESSEEG | — | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — share above chance: not computed — Status: Published | [On this site →](https://bci.report/topics/clinical-groups/#iesseeg) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-iesseeg) | | [Kim beta-range SSVEP · 40 targets, 14.0–21.8 Hz (Figshare 28806815) ↗](https://doi.org/10.6084/m9.figshare.28806815.v3) — MOABB: Kim2025BetaRange · DOI: 10.1038/s41597-025-06032-2 | 40-target SSVEP speller, 14.0 to 21.8 Hz — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | Held: models whose output depends on amplitude, because amplitude units, channel gains and the saved reference history are not established. [On this site →](https://bci.report/datasets/kim2025-beta-range/) · [The case →](https://bci.report/topics/data-quality/#kim2025-beta-range) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-kim2025-beta-range) | | [Longitudinal ERP dataset (RSVP) ↗](https://doi.org/10.6084/m9.figshare.27201003.v1) — Also: A longitudinal EEG dataset of event-related potential, figshare 27201003, Longitudinal ERP dataset, Zhang et al. (2025) · MOABB: Zhang2025 · DOI: 10.1038/s41597-025-05378-x | RSVP face-target ERP, the same people at nominal Day 1, 7, 80 and 200 visits — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/longitudinal-rsvp/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-longitudinal-rsvp) | | [LTRSVP · EEG Signals from an RSVP Task ↗](https://physionet.org/content/ltrsvp/1.0.0/) — Also: LTRSVP, EEG Signals from an RSVP Task, PhysioNet ltrsvp, Matran-Fernandez & Poli (2017) | P300 target images in rapid serial visual presentation at three rates — as this site’s page states it | Open Data Commons Attribution License 1.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/ltrsvp/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ltrsvp) | | [MIND · multimodal fNIRS-EEG unilateral limb motor imagery (ScienceDB 34326) ↗](https://doi.org/10.57760/sciencedb.34326) — MOABB: MIND2026 · DOI: 10.57760/sciencedb.34326 | — | CC BY 4.0, CC BY-NC 4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 3: original terms — gap at step 3: pinned version — gap at step 4: chance level — share above chance: not computed — Status: Published | [The case →](https://bci.report/topics/data-quality/#mind2026) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-mind2026) | | [Mobile BCI dataset (SSVEP and ERP paradigms) ↗](https://doi.org/10.82901/nemar.nm000125) — Also: Mobile BCI dataset, Lee2021, Lee2021Mobile, NEMAR nm000201, NEMAR nm000125, OSF r7s9b · MOABB: Lee2021Mobile_ERP, Lee2021Mobile_SSVEP · DOI: 10.1038/s41597-021-01094-4 · EEGDash: EEGDash/nm000125, EEGDash/nm000201 | SSVEP and ERP while standing, walking and running — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/mobile-bci/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-mobile-bci) | | [NeBULA · EEG-EMG standardized reaching task (Figshare 27301629) ↗](https://doi.org/10.6084/m9.figshare.27301629.v1) — MOABB: NeBULA2025 · DOI: 10.1038/s41597-025-05042-4 | — | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — Status: Score held | Held: with one person, every score is that person’s. [The case →](https://bci.report/topics/data-quality/#nebula) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-nebula) | | [MC_RTT (Neural Latents Benchmark) ↗](https://doi.org/10.48324/dandi.000129/0.241017.1444) — Also: MC_RTT, Neural Latents Benchmark MC_RTT, DANDI 000129, macaque motor cortex spiking activity during self-paced reaching, O'Doherty et al. random target task | One macaque’s motor-cortex spiking during self-paced reaching: finger velocity — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — Status: Published | [On this site →](https://bci.report/datasets/nlb-mc-rtt/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-nlb-mc-rtt) | | [OpenBMI motor imagery (Lee et al. 2019) ↗](https://doi.org/10.5524/100542) — Also: OpenBMI, Lee2019, GigaDB 100542, EEG Dataset and OpenBMI Toolbox for Three BCI Paradigms, Lee et al. (2019) · MOABB: Lee2019_MI, Lee2019_ERP, Lee2019_SSVEP · DOI: 10.1093/gigascience/giz002 | Motor imagery: left or right hand across two sessions, and imagery or rest — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | eeg-fm-masking lists it in its pretraining corpus; which release of it it used is not established. [On this site →](https://bci.report/datasets/openbmi/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-openbmi) | | [Orthosis error-detection dataset (Kueper et al. 2024) ↗](https://doi.org/10.5281/zenodo.8345429) — Also: EEG and EMG dataset for the detection of errors introduced by an active orthosis device, IJCAI'23 CC6 IntEr-HRI competition dataset, Zenodo 8345429, Kueper et al. (2024) | Injected orthosis device errors against standard device events, EEG only: one cohort in three configurations — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/orthosis-error-detection/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-orthosis-error-detection) | | [Phantom EEG with motion, muscle and eye artifacts ↗](https://doi.org/10.18112/openneuro.ds004784.v1.0.4) — Also: Phantom EEG Dataset with Motion, Muscle, and Eye Artifacts and Example Scripts, OpenNeuro ds004784 · EEGDash: EEGDash/ds004784 | Physical head phantom with motion, muscle and eye artifacts — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 4 of 5 · Run — Status: Published | [On this site →](https://bci.report/datasets/phantom/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-phantom) | | [REFED emotion dataset (EEG-fNIRS) ↗](https://huggingface.co/datasets?search=REFED-dataset) — Also: REFED, A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset, REFED2025/REFED-dataset, Ning et al. (2025) | Self-reported valence and arousal while watching videos: a stored-clip diagnostic — as this site’s page states it | CC BY-NC-SA 4.0 — this site’s rights record | Step 4 of 5 · Run — Status: Published | [On this site →](https://bci.report/datasets/refed/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-refed) | | [Shin2022 · closed-loop motor imagery EEG (Figshare 20383716) ↗](https://doi.org/10.6084/m9.figshare.20383716.v1) — MOABB: Shin2022 · DOI: 10.3389/fnhum.2022.951591 | Closed-loop cursor control: the instructed right or left target, dry electrodes — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — share above chance: not computed — Status: Published | MOABB’s catalog lists 2,640 trials and 11 runs for it, which differ from what its files hold: see [the case on the data-quality page](https://bci.report/topics/data-quality/#shin2022). [On this site →](https://bci.report/datasets/shin2022/) · [The case →](https://bci.report/topics/data-quality/#shin2022) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-shin2022) | | [Stieger longitudinal BCI · one-person pilot ↗](https://doi.org/10.6084/m9.figshare.13123148.v1) — MOABB: Stieger2021 · DOI: 10.1038/s41597-021-00883-1 | — | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — Status: Status only | [The case →](https://bci.report/topics/model-adaptation/#cross-session) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-stieger-longitudinal) | | [Cattan PC/VR P300 dataset ↗](https://doi.org/10.5281/zenodo.2605205) — Also: Dataset of an EEG-based BCI experiment in Virtual Reality and on a Personal Computer, VR.EEG.2018-GIPSA, Zenodo 2605205, arXiv:1903.11297 · MOABB: Cattan2019_VR · DOI: 10.5281/zenodo.2605204 | P300 on a PC screen and in a VR headset — as this site’s page states it | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/vr-pc-p300/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-vr-pc-p300) | | [WBCIC-SHU motor imagery dataset ↗](https://doi.org/10.25452/figshare.plus.22671172.v5) — Also: WBCIC-SHU, WBCIC-SHU Motor Imagery Dataset, WBCIC_SHU Motor Imagery dataset, Figshare+ 22671172, A multi-day and high-quality EEG dataset for motor imagery brain-computer interface, Yang et al. (2025) · MOABB: Yang2025 · DOI: 10.1038/s41597-025-04826-y | Motor imagery across three recording sessions: two-class and three-class cohorts, kept apart — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/wbcic-shu/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-wbcic-shu) | | [Wearable SSVEP BCI dataset (dry and wet electrodes) ↗](https://doi.org/10.6084/m9.figshare.13560281.v4) — Also: Wearable SSVEP BCI Dataset, An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces, figshare 13560281, Zhu et al. (2021) | 12-target SSVEP, dry and wet electrodes — as this site’s page states it | CC BY 4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 3: pinned version — gap at step 4: chance level — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/wearable-ssvep-102/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-wearable-ssvep-102) | | [YSU asynchronous SSVEP-BCI dataset ↗](https://doi.org/10.6084/m9.figshare.24906300.v3) — Also: YSU-an EEG dataset for studying asynchronous SSVEP-BCI, figshare 24906300, An EEG dataset for studying asynchronous steady-state visual evoked potential (SSVEP) based brain computer interfaces | Asynchronous SSVEP, control and non-control states — as this site’s page states it | CC-BY-4.0 — this site’s rights record | Step 4 of 5 · Run — gap at step 4: chance level — share above chance: not computed — Status: Published | [On this site →](https://bci.report/datasets/ysu-async-ssvep/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ysu-async-ssvep) | | [BNCI2014-001 ↗](https://bnci-horizon-2020.eu/database/data-sets/001-2014) — Also: BCI Competition IV 2a · MOABB: BNCI2014_001 · DOI: 10.3389/fnins.2012.00055 | — | CC BY-ND 4.0 — this site’s rights record | Step 3 of 5 · Traced — gap at step 3: consent or ethics statement — gap at step 3: pinned version — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bnci2014-001) | | [BNCI2014-004 ↗](https://bnci-horizon-2020.eu/database/data-sets/004-2014) — MOABB: BNCI2014_004 · DOI: 10.1109/TNSRE.2007.906956 | — | CC BY-ND 4.0 — this site’s rights record | Step 3 of 5 · Traced — gap at step 2: files opened — gap at step 3: consent or ethics statement — gap at step 3: pinned version — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bnci2014-004) | | [BNCI2014-008 ↗](https://bnci-horizon-2020.eu/database/data-sets/008-2014) — MOABB: BNCI2014_008 · DOI: 10.3389/fnhum.2013.00732 | — | CC BY-NC-ND 4.0 — this site’s rights record | Step 3 of 5 · Traced — gap at step 3: pinned version — Status: Listed | REVE Base and REVE Large list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bnci2014-008) | | [BNCI2014-009 ↗](https://bnci-horizon-2020.eu/database/data-sets/009-2014) — MOABB: BNCI2014_009 · DOI: 10.1088/1741-2560/11/3/035008 | — | CC BY-NC-ND 4.0 — this site’s rights record | Step 3 of 5 · Traced — gap at step 3: consent or ethics statement — gap at step 3: pinned version — Status: Listed | REVE Base and REVE Large list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-bnci2014-009) | | [OpenNeuro ds004902 · paired sleep-condition recordings ↗](https://openneuro.org/datasets/ds004902/versions/1.0.8) — EEGDash: EEGDash/ds004902 | — | CC0-1.0 — this site’s rights record | Step 3 of 5 · Traced — gap at step 3: consent or ethics statement — Status: Status only | eeg-fm-masking lists it in its pretraining corpus; which release of it it used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ds004902) | | [OpenNeuro ds007420 · multi-distance fNIRS ↗](https://doi.org/10.18112/openneuro.ds007420.v1.0.5) — Also: ds007420, Open-access Multi-Distance fNIRS Dataset Generated During Ball-Squeezing and Motion-Artifact Induction Tasks, Zhou et al. multi-distance fNIRS · EEGDash: EEGDash/ds007420 | Human fNIRS recordings: a numerical coverage check, no decoding — as this site’s page states it | CC0-1.0 — this site’s rights record | Step 3 of 5 · Traced — Status: Published | [On this site →](https://bci.report/datasets/ds007420/) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-ds007420) | | [FALCON H2 · human intracortical handwriting ↗](https://doi.org/10.48324/dandi.000950/0.241029.1403) | — | CC BY 4.0 — this site’s rights record | Step 3 of 5 · Traced — Status: Status only | [The case →](https://bci.report/topics/beyond-scalp-eeg/#falcon-h2) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-falcon-h2) | | [physionet-eegmmidb-1.0.0 ↗](https://physionet.org/content/eegmmidb/1.0.0/) — MOABB: PhysionetMI · DOI: 10.13026/C28G6P | — | Open Data Commons Attribution License 1.0 — this site’s rights record | Step 3 of 5 · Traced — gap at step 2: files opened — gap at step 3: consent or ethics statement — Status: Listed | LaBraM lists it in its pretraining corpus; which release of it it used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-physionet-eegmmidb) | | [Rabut et al. · human functional ultrasound through a cranial window ↗](https://doi.org/10.22002/f3y3k-em558) | — | CC BY-NC 4.0 — this site’s rights record | Step 3 of 5 · Traced — Status: Status only | [The case →](https://bci.report/topics/beyond-scalp-eeg/#rabut-fus) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-rabut-fus) | | [L-FAME ↗](https://huggingface.co/datasets?search=L-FAME) — Hugging Face: L-FAME-Dataset-Benchmark/L-FAME | — | CC-BY-NC-4.0 — this site’s rights record | Step 2 of 5 · Read — Status: Status only | Listed because the register records what was checked; it has no model result on this site. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-lfame) | | [Synchronized multimodal meditation and rest dataset (Zenodo) ↗](https://zenodo.org/records/17735692) | — | None recorded | Step 2 of 5 · Read — gap at step 1: licence as stated — Status: Status only | [The case →](https://bci.report/topics/pilot-studies/#status-only) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-meditation-rest) | | [stew-monster-processed ↗](https://ieee-dataport.org/open-access/simultaneous-task-eeg-workload-dataset) — Hugging Face: monster-monash/STEW | — | Mirror declares CC BY 4.0; exact upstream DataPort license not independently verified — this site’s rights record | Step 2 of 5 · Read — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-stew-processed) | | [ZuCo 2.0 via NEMAR nm000279 ↗](https://doi.org/10.82901/nemar.nm000279.v1.0.0) | — | CC-BY-4.0 — this site’s rights record | Step 2 of 5 · Read — Status: Status only | Held: in one converted run, some fixation times fall outside the recording. [The case →](https://bci.report/topics/data-quality/#zuco2) · [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-site-zuco2) | | **MOABB catalog** · 136 rows — One row per source record in MOABB, with paradigm, people, channels and sampling rate as MOABB lists them. Records this site already lists are merged into its rows. | | | | | | [AguileraRodriguez2025 ↗](https://doi.org/10.1038/s41597-025-05926-5) — DOI: 10.1038/s41597-025-05926-5 | Imagery · 15 people · 24 channels · 500 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0001) | | [Alawieh2025 ↗](https://doi.org/10.5281/zenodo.15454354) — DOI: 10.5281/zenodo.15454354 | Imagery · 27 people · 35 channels · 512 Hz — as MOABB lists it | cc-by-4.0 — the deposit’s record | Step 1 of 5 · Listed — Status: Listed | Held: MOABB loads one offline session, and the clinical group has two participants. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0002) | | [AlexMI ↗](https://zenodo.org/record/806023) — DOI: 10.5281/zenodo.806022 | Imagery · 8 people · 16 channels · 512 Hz — as MOABB lists it | CC-BY-SA-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0003) | | [BCIComp2020 ↗](https://osf.io/pq7vb/) — MOABB: BCIComp2020UpperLimb, BCIComp2020WalkingERP · DOI: 10.3389/fnhum.2022.898300 | BCIComp2020UpperLimb: Imagery · 15 people · 60 channels · 250 Hz BCIComp2020WalkingERP: P300 / ERP · 15 people · 56 channels · 100 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0004) | | [BI2012 ↗](https://doi.org/10.5281/zenodo.2649006) — DOI: 10.5281/zenodo.2649006 | P300 / ERP · 25 people · 16 channels · 128 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | LaBraM and REVE list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0005) | | [BI2013a ↗](https://doi.org/10.5281/zenodo.2669187) — DOI: 10.5281/zenodo.1494163, 10.5281/zenodo.2669187 | P300 / ERP · 24 people · 16 channels · 512 Hz — as MOABB lists it | Creative Commons Attribution 1.0 Generic — the Zenodo record — Zenodo record 2669187, the latest version under concept DOI 10.5281/zenodo.1494163; licence identifier on Zenodo: “cc-by” | Step 3 of 5 · Traced — gap at step 2: files opened — Status: Planned | LaBraM and REVE list it in their pretraining corpus; which release of it they used is not established.; Planned for a fixed-protocol run here. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0006) | | [BI2014a ↗](https://doi.org/10.5281/zenodo.3266223) — DOI: 10.5281/zenodo.3266223 | P300 / ERP · 64 people · 16 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | LaBraM and REVE list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0007) | | [BI2014b ↗](https://doi.org/10.5281/zenodo.3267301) — DOI: 10.5281/zenodo.3267301 | P300 / ERP · 38 people · 32 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | LaBraM and REVE list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0008) | | [BI2015a ↗](https://doi.org/10.5281/zenodo.3266930) — DOI: 10.5281/zenodo.3266930 | P300 / ERP · 43 people · 32 channels · 512 Hz — as MOABB lists it | cc-by-4.0 — the deposit’s record | Step 1 of 5 · Listed — Status: Listed | LaBraM and REVE list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0009) | | [BI2015b ↗](https://doi.org/10.5281/zenodo.3267307) — DOI: 10.5281/zenodo.3267307 | P300 / ERP · 44 people · 32 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | LaBraM and REVE list it in their pretraining corpus; which release of it they used is not established. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0010) | | [BNCI2003_004 ↗](https://doi.org/10.1109/TBME.2004.827088) — DOI: 10.1109/TBME.2004.827088 | Imagery · 5 people · 118 channels · 100 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0011) | | [BNCI2014_002 ↗](https://doi.org/10.1515/bmt-2014-0117) — DOI: 10.1515/bmt-2014-0117 | Imagery · 14 people · 15 channels · 512 Hz — as MOABB lists it | CC-BY-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0013) | | [BNCI2015_003 ↗](https://doi.org/10.1016/j.neulet.2009.06.045) — DOI: 10.1016/j.neulet.2009.06.045 | P300 / ERP · 10 people · 8 channels · 256 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0018) | | [BNCI2015_004 ↗](https://bnci-horizon-2020.eu/database/data-sets) — DOI: 10.1371/journal.pone.0123727 | Imagery · 9 people · 30 channels · 256 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0019) | | [BNCI2015_006 ↗](https://github.com/bbci/bbci_public/blob/master/doc/index.markdown) — DOI: 10.1088/1741-2560/11/2/026009 | P300 / ERP · 11 people · 64 channels · 200 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0020) | | [BNCI2015_007 ↗](https://doi.org/10.1088/1741-2560/9/4/045006) — DOI: 10.1088/1741-2560/9/4/045006 | P300 / ERP · 16 people · 63 channels · 100 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0021) | | [BNCI2015_008 ↗](https://github.com/bbci/bbci_public/blob/master/doc/index.markdown) — DOI: 10.1088/1741-2560/8/6/066003 | P300 / ERP · 13 people · 63 channels · 250 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0022) | | [BNCI2015_009 ↗](https://doi.org/10.3389/fnins.2011.00112) — DOI: 10.3389/fnins.2011.00112 | P300 / ERP · 21 people · 62 channels · 1,000 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0023) | | [BNCI2015_010 ↗](https://doi.org/10.1016/j.clinph.2012.12.050) — DOI: 10.1016/j.clinph.2012.12.050 | P300 / ERP · 12 people · 63 channels · 200 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0024) | | [BNCI2015_012 ↗](https://doi.org/10.3389/fnins.2011.00099) — DOI: 10.3389/fnins.2011.00099 | P300 / ERP · 10 people · 63 channels · 250 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0025) | | [BNCI2015_013 ↗](https://doi.org/10.1109/TNSRE.2010.2053387) — DOI: 10.1109/TNSRE.2010.2053387 | P300 / ERP · 6 people · 64 channels · 512 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0026) | | [BNCI2016_002 ↗](https://doi.org/10.1088/1741-2560/8/5/056001) — DOI: 10.1088/1741-2560/8/5/056001 | P300 / ERP · 15 people · 69 channels · 200 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0027) | | [BNCI2019_001 ↗](https://doi.org/10.5281/zenodo.2222268) — DOI: 10.1038/s41598-019-43594-9 | Imagery · 10 people · 64 channels · 256 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0028) | | [BNCI2020_001 ↗](https://bnci-horizon-2020.eu/database/data-sets) — DOI: 10.3389/fnins.2020.00849 | Imagery · 45 people · 64 channels · 256 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0029) | | [BNCI2020_002 ↗](http://bnci-horizon-2020.eu/database/data-sets) — DOI: 10.3389/fnins.2020.591777 | P300 / ERP · 18 people · 32 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0030) | | [BNCI2022_001 ↗](https://doi.org/10.1109/TAFFC.2021.3059688) — DOI: 10.1109/TAFFC.2021.3059688 | Imagery · 13 people · 67 channels · 256 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0031) | | [BNCI2024_001 ↗](https://bnci-horizon-2020.eu/database/data-sets) — DOI: 10.1016/j.compbiomed.2024.109132 | cognitive · 20 people · 64 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0032) | | [BNCI2025_001 ↗](https://github.com/rkobler/eyeartifactcorrection) — DOI: 10.1088/1741-2552/ada0ea | Imagery · 20 people · 71 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0033) | | [BNCI2025_002 ↗](https://github.com/sccn/labstreaminglayer) — DOI: 10.1088/1741-2552/ac689f | Imagery · 10 people · 64 channels · 200 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0034) | | [Batista2022 ↗](https://zenodo.org/records/7664069) — DOI: 10.5281/zenodo.7664068 | Imagery · 20 people · 38 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0035) | | [Beetl2021_A ↗](https://beetl.ai/data) — DOI: 10.48550/arXiv.2202.12950 | Imagery · 3 people · 63 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0036) | | [Beetl2021_B ↗](https://beetl.ai/data) | Imagery · 2 people · 32 channels · 200 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0037) | | [Brandl2020 ↗](https://depositonce.tu-berlin.de/handle/11303/10934.2) — DOI: 10.3389/fnins.2020.566147 | Imagery · 16 people · 63 channels · 1,000 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0038) | | [Brodu2012 ↗](https://openvibe.inria.fr/private/datasets/dataset-1/) — DOI: 10.1016/j.neucom.2011.10.010 | Imagery · 1 person · 11 channels · 512 Hz — as MOABB lists it | “free for research use” — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0039) | | [Castillos ↗](https://zenodo.org/record/8255618) — MOABB: CastillosBurstVEP100, CastillosBurstVEP40, CastillosCVEP100, CastillosCVEP40 · DOI: 10.1016/j.neuroimage.2023.120446 | CastillosBurstVEP100: c-VEP · 12 people · 28 channels · 500 Hz CastillosBurstVEP40: c-VEP · 12 people · 28 channels · 500 Hz CastillosCVEP100: c-VEP · 12 people · 28 channels · 500 Hz CastillosCVEP40: c-VEP · 12 people · 28 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0040) | | [Cattan2019_PHMD ↗](https://doi.org/10.5281/zenodo.2617084) — DOI: 10.5281/zenodo.2617084 | Resting State · 12 people · 16 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0041) | | [Chailloux2020 ↗](https://openneuro.org/datasets/ds003190) — DOI: 10.3390/s20247198 · EEGDash: EEGDash/ds003190 | P300 / ERP · 19 people · 8 channels · 256 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0043) | | [Chang2025 ↗](https://figshare.com/articles/dataset/28831730) — DOI: 10.1038/s41597-025-06147-6 | Imagery · 28 people · 59 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0044) | | [Chen2017SingleFlicker ↗](https://zenodo.org/records/580485) — DOI: 10.1371/journal.pone.0178385 | SSVEP · 12 people · 32 channels · 512 Hz — as MOABB lists it | CC BY 4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0045) | | [Cho2017 ↗](http://dx.doi.org/10.5524/100295) — DOI: 10.5524/100295 | Imagery · 52 people · 68 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0046) | | [Daly2020 ↗](https://openneuro.org/datasets/ds002720) — DOI: 10.18112/openneuro.ds002720.v1.0.1 · EEGDash: EEGDash/ds002720 | Imagery · 18 people · 19 channels · 1,000 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0047) | | [Dong2023 ↗](https://zenodo.org/records/18847318) — DOI: 10.26599/BSA.2023.9050020 | SSVEP · 59 people · 8 channels · 250 Hz — as MOABB lists it | CC BY-NC 4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0049) | | [Dreyer2023 ↗](https://doi.org/10.5281/zenodo.8089820) — MOABB: Dreyer2023A, Dreyer2023B, Dreyer2023C · DOI: 10.1038/s41597-023-02445-z | Dreyer2023: Imagery · 87 people · 32 channels · 512 Hz Dreyer2023A: Imagery · 60 people · 32 channels · 512 Hz Dreyer2023B: Imagery · 21 people · 32 channels · 512 Hz Dreyer2023C: Imagery · 6 people · 32 channels · 512 Hz — as MOABB lists it | cc-by-4.0 — the deposit’s record | Step 1 of 5 · Listed — Status: Listed | Held: three releases form one family; selection and filter differences are unchecked. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0050) | | [EPFLP300 ↗](https://doi.org/10.1016/j.jneumeth.2007.03.005) — DOI: 10.1016/j.jneumeth.2007.03.005 | P300 / ERP · 8 people · 34 channels · 2,048 Hz — as MOABB lists it | Licence not yet read from the host | Step 1 of 5 · Listed — gap at step 1: licence as stated — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0051) | | [ErpCore2021 ↗](https://doi.org/10.18115/D5JW4R) — MOABB: ErpCore2021_ERN, ErpCore2021_LRP, ErpCore2021_MMN, ErpCore2021_N170, ErpCore2021_N2pc, ErpCore2021_N400, ErpCore2021_P3 · DOI: 10.1016/j.neuroimage.2020.117465 | ErpCore2021: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_ERN: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_LRP: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_MMN: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_N170: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_N2pc: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_N400: P300 / ERP · 40 people · 33 channels · 1,024 Hz ErpCore2021_P3: P300 / ERP · 40 people · 33 channels · 1,024 Hz — as MOABB lists it | CC BY 4.0, CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0052) | | [Farabbi2020 ↗](https://doi.org/10.5281/zenodo.5882500) — DOI: 10.5281/zenodo.5882500 | Imagery · 12 people · 35 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0053) | | [Forenzo2023 ↗](https://kilthub.cmu.edu/articles/dataset/23677098) — DOI: 10.1109/TBME.2023.3298957 | Imagery · 25 people · 64 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0054) | | [Gao2026 ↗](https://doi.org/10.6084/m9.figshare.30227503.v1) — DOI: 10.1038/s41597-025-06512-5 | Imagery · 22 people · 32 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0055) | | [GrosseWentrup2009 ↗](https://doi.org/10.1109/TBME.2008.2009768) — DOI: 10.1109/TBME.2008.2009768 | Imagery · 10 people · 128 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0056) | | [GuttmannFlury2025 ↗](https://doi.org/10.1038/s41597-025-04861-9) — MOABB: GuttmannFlury2025_ME, GuttmannFlury2025_MI, GuttmannFlury2025_P300, GuttmannFlury2025_SSVEP · DOI: 10.1038/s41597-025-04861-9 | GuttmannFlury2025_ME: Imagery · 31 people · 63 channels · 1,000 Hz GuttmannFlury2025_MI: Imagery · 31 people · 63 channels · 1,000 Hz GuttmannFlury2025_P300: P300 / ERP · 31 people · 63 channels · 1,000 Hz GuttmannFlury2025_SSVEP: SSVEP · 31 people · 63 channels · 1,000 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0057) | | [Han2024Fatigue ↗](https://zenodo.org/records/10507229) — DOI: 10.1109/TNSRE.2024.3380635 | SSVEP · 24 people · 64 channels · 1,000 Hz — as MOABB lists it | CC BY 4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0058) | | [Han2026 ↗](https://openneuro.org/datasets/ds007327) — DOI: 10.18112/openneuro.ds007327.v1.1.0 | Imagery · 35 people · 88 channels · 1,000 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0059) | | [HefmiIch2025 ↗](https://figshare.com/articles/dataset/28955456) — DOI: 10.1038/s41597-025-06100-7 | Imagery · 37 people · 32 channels · 256 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0060) | | [Hinss2021 ↗](https://doi.org/10.5281/zenodo.6874128) — DOI: 10.1038/s41597-022-01898-y | Resting State · 15 people · 62 channels · 500 Hz — as MOABB lists it | cc-by-sa-4.0 — the deposit’s record | Step 1 of 5 · Listed — Status: Listed | Held: two deposits differ in people, sessions and licence; the version is not fixed yet. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0061) | | [Huebner2017 ↗](http://doi.org/10.5281/zenodo.192684) — DOI: 10.1371/journal.pone.0175856 | P300 / ERP · 13 people · 37 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0062) | | [Huebner2018 ↗](https://zenodo.org/record/5831879) — DOI: 10.1109/MCI.2018.2807039 | P300 / ERP · 12 people · 37 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0063) | | [IMUMIA2026 ↗](https://doi.org/10.5281/zenodo.20421767) — DOI: 10.5281/zenodo.20421767 | Imagery · 5 people · 69 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0064) | | [Iwama2023 ↗](https://openneuro.org/datasets/ds004444) — DOI: 10.1038/s41597-023-02260-6 · EEGDash: EEGDash/ds004444 | Imagery · 30 people · 129 channels · 1,000 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0065) | | [Jeong2020 ↗](https://zenodo.org/records/19021436) — DOI: 10.1093/gigascience/giaa098 | Imagery · 25 people · 71 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0066) | | [Jia2019 ↗](https://doi.org/10.6084/m9.figshare.7636301) — DOI: 10.6084/m9.figshare.7636301 | Imagery · 15 people · 63 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0067) | | [Kalunga2016 ↗](https://zenodo.org/record/2392979) — DOI: 10.1016/j.neucom.2016.01.007 | SSVEP · 12 people · 8 channels · 256 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0068) | | [Kaneshiro2015 ↗](https://purl.stanford.edu/bq914sc3730) — DOI: 10.1371/journal.pone.0135697 | Imagery · 10 people · 124 channels · 62.5 Hz — as MOABB lists it | CC-BY-3.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0069) | | [Kaya2018 ↗](https://figshare.com/collections/A_large_electroencephalographic_motor_imagery_dataset_for_electroencephalographic_brain_computer_interfaces/3917698) — DOI: 10.1038/sdata.2018.211 | Imagery · 7 people · 19 channels · 200 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0070) | | [Kodera2023 ↗](https://doi.org/10.5281/zenodo.7893846) — DOI: 10.5281/zenodo.7893846 | Imagery · 29 people · 9 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0072) | | [Kojima2024A ↗](https://doi.org/10.7910/DVN/MQOVEY) — DOI: 10.1371/journal.pone.0303565 | P300 / ERP · 11 people · 66 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0073) | | [Kojima2024B ↗](https://doi.org/10.7910/DVN/1UJDV6) — DOI: 10.3389/fnhum.2024.1461960 | P300 / ERP · 15 people · 66 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0074) | | [Kueper2024 ↗](https://doi.org/10.5281/zenodo.10229480) — DOI: 10.1038/s41598-024-65910-8 | Imagery · 8 people · 67 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0075) | | [Kumar2024 ↗](https://zenodo.org/records/10694880) — DOI: 10.1093/pnasnexus/pgae076 | Imagery · 18 people · 25 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0076) | | [Lee2022 ↗](https://openneuro.org/datasets/ds004022) — DOI: 10.18112/openneuro.ds004022.v1.0.0 · EEGDash: EEGDash/ds004022 | Imagery · 7 people · 18 channels · 500 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0079) | | [Lee2024 ↗](https://github.com/jml226/Home-Appliance-Control-Dataset) — MOABB: Lee2024_AC, Lee2024_BS, Lee2024_DL, Lee2024_EL, Lee2024_TV · DOI: 10.3389/fnhum.2024.1320457 | Lee2024_AC: P300 / ERP · 10 people · 25 channels · 500 Hz Lee2024_BS: P300 / ERP · 14 people · 31 channels · 500 Hz Lee2024_DL: P300 / ERP · 15 people · 31 channels · 500 Hz Lee2024_EL: P300 / ERP · 15 people · 31 channels · 500 Hz Lee2024_TV: P300 / ERP · 30 people · 31 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0080) | | [Leelakittisin2025 ↗](https://doi.org/10.5281/zenodo.20348444) — DOI: 10.5281/zenodo.20348444 | Imagery · 22 people · 63 channels · 1,200 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0081) | | [Leeuwis2021 ↗](https://doi.org/10.34894/Z7ZVOD) — DOI: 10.34894/Z7ZVOD | Imagery · 55 people · 16 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0082) | | [Lenaig2026 ↗](https://doi.org/10.5281/zenodo.21156618) — DOI: 10.1109/BCI69045.2026.11435072 | SSVEP · 24 people · 27 channels · 500 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0083) | | [Lioi2020 ↗](https://openneuro.org/datasets/ds002336) — MOABB: Lioi2020_XP1, Lioi2020_XP2 · DOI: 10.1038/s41597-020-0498-3 · EEGDash: EEGDash/ds002336 | Lioi2020_XP1: Imagery · 10 people · 64 channels · 5,000 Hz Lioi2020_XP2: Imagery · 17 people · 64 channels · 5,000 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0084) | | [Liu2024 ↗](https://doi.org/10.6084/m9.figshare.21679035.v5) — DOI: 10.1038/s41597-023-02787-8 | Imagery · 47 people · 31 channels · 500 Hz — as MOABB lists it | CC BY 4.0 — the deposit’s record | Step 1 of 5 · Listed — Status: Listed | Held: released IDs include duplicated or reprocessed recordings; group-level deduplication is pending. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0087) | | [Liu2025 ↗](https://zenodo.org/records/18987384) — DOI: 10.1038/s41597-025-04618-4 | Imagery · 27 people · 65 channels · 1,000 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0088) | | [MAMEM1 ↗](https://dx.doi.org/10.6084/m9.figshare.2068677.v1) — DOI: 10.6084/m9.figshare.2068677.v1 | SSVEP · 11 people · 256 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0089) | | [MAMEM2 ↗](https://figshare.com/articles/dataset/3153409) — DOI: 10.48550/arXiv.1602.00904 | SSVEP · 11 people · 256 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0090) | | [MAMEM3 ↗](https://doi.org/10.6084/m9.figshare.3413851.v3) — DOI: 10.6084/m9.figshare.3413851.v3 | SSVEP · 11 people · 14 channels · 128 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0091) | | [MIBMPI2024 ↗](https://zenodo.org/records/13626922) — DOI: 10.1007/s00521-024-10917-5 | Imagery · 8 people · 13 channels · 128 Hz — as MOABB lists it | CC-BY-NC-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0092) | | [MILimbEEG ↗](https://doi.org/10.17632/x8psbz3f6x.2) — DOI: 10.1016/j.dib.2023.109540 | Imagery · 60 people · 16 channels · 125 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0093) | | [MIMED2024 ↗](https://doi.org/10.17632/zs25xxjkm9.3) — DOI: 10.1016/j.dib.2024.110833 | Imagery · 30 people · 14 channels · 128 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0094) | | [MOVING2024 ↗](https://doi.org/10.5281/zenodo.12804784) — DOI: 10.3390/s24165207 | Imagery · 11 people · 35 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0096) | | [Ma2020 ↗](https://doi.org/10.7910/DVN/RBN3XG) — DOI: 10.1038/s41597-020-0535-2 | Imagery · 25 people · 62 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0097) | | [Ma2022 ↗](https://nemar.org/dataexplorer/detail?dataset_id=nm000288) — DOI: 10.1038/s41597-022-01647-1 | Imagery · 25 people · 32 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0098) | | [Mainsah2025 ↗](https://physionet.org/content/bigp3bci/1.0.0/) — MOABB: Mainsah2025_A, Mainsah2025_B, Mainsah2025_C, Mainsah2025_D, Mainsah2025_E, Mainsah2025_F, Mainsah2025_G, Mainsah2025_H, Mainsah2025_I, Mainsah2025_J, Mainsah2025_K, Mainsah2025_L, Mainsah2025_M, Mainsah2025_N, Mainsah2025_O, Mainsah2025_P, Mainsah2025_Q, Mainsah2025_R, Mainsah2025_S1, Mainsah2025_S2 · DOI: 10.13026/0byy-ry86 | Mainsah2025_A: P300 / ERP · 13 people · 32 channels · 256 Hz Mainsah2025_B: P300 / ERP · 19 people · 16 channels · 256 Hz Mainsah2025_C: P300 / ERP · 19 people · 32 channels · 256 Hz Mainsah2025_D: P300 / ERP · 17 people · 32 channels · 256 Hz Mainsah2025_E: P300 / ERP · 8 people · 16 channels · 256 Hz Mainsah2025_F: P300 / ERP · 10 people · 16 channels · 256 Hz Mainsah2025_G: P300 / ERP · 20 people · 16 channels · 256 Hz Mainsah2025_H: P300 / ERP · 16 people · 16 channels · 256 Hz Mainsah2025_I: P300 / ERP · 13 people · 16 channels · 256 Hz Mainsah2025_J: P300 / ERP · 20 people · 16 channels · 256 Hz Mainsah2025_K: P300 / ERP · 5 people · 16 channels · 256 Hz Mainsah2025_L: P300 / ERP · 11 people · 16 channels · 256 Hz Mainsah2025_M: P300 / ERP · 21 people · 16 channels · 256 Hz Mainsah2025_N: P300 / ERP · 8 people · 16 channels · 256 Hz Mainsah2025_O: P300 / ERP · 18 people · 32 channels · 256 Hz Mainsah2025_P: P300 / ERP · 19 people · 32 channels · 256 Hz Mainsah2025_Q: P300 / ERP · 36 people · 32 channels · 256 Hz Mainsah2025_R: P300 / ERP · 20 people · 32 channels · 256 Hz Mainsah2025_S1: P300 / ERP · 10 people · 32 channels · 256 Hz Mainsah2025_S2: P300 / ERP · 24 people · 32 channels · 256 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0099) | | [MartinezCagigal2023Checker ↗](https://doi.org/10.71569/7c67-v596) — DOI: 10.71569/7c67-v596 | c-VEP · 16 people · 16 channels · 256 Hz — as MOABB lists it | CC-BY-NC-SA-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0100) | | [MartinezCagigal2023Pary ↗](https://doi.org/10.71569/025s-eq10) — DOI: 10.71569/025s-eq10 | c-VEP · 16 people · 16 channels · 256 Hz — as MOABB lists it | CC-BY-NC-SA-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0101) | | [MartinezPeon2024 ↗](https://doi.org/10.6084/m9.figshare.25773342) — DOI: 10.6084/m9.figshare.25773342.v1 | Imagery · 10 people · 14 channels · 128 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0102) | | [MartinezPeon2025 ↗](https://data.mendeley.com/datasets/msgzn862ns/1) — DOI: 10.17632/msgzn862ns.1 | Imagery · 50 people · 20 channels · 500 Hz — as MOABB lists it | CC-BY-NC-SA-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0103) | | [NETBCI2026 ↗](https://doi.org/10.57745/RBJRC7) — DOI: 10.1038/s41597-026-08237-5 | Imagery · 19 people · 74 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0104) | | [Nakanishi2015 ↗](https://github.com/mnakanishi/12JFPM_SSVEP/raw/master/data/) — DOI: 10.1371/journal.pone.0140703 | SSVEP · 10 people · 8 channels · 256 Hz — as MOABB lists it | Licence not yet read from the host | Step 1 of 5 · Listed — gap at step 1: licence as stated — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0105) | | [NeuroTUMBCI2025 ↗](https://doi.org/10.5281/zenodo.18087806) — DOI: 10.5281/zenodo.18087806 | Imagery · 2 people · 24 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0107) | | [Nguyen2017 ↗](https://zenodo.org/records/19502794) — MOABB: Nguyen2017_L, Nguyen2017_S, Nguyen2017_SL, Nguyen2017_V · DOI: 10.1088/1741-2552/aa8235 | Nguyen2017_L: Imagery · 6 people · 64 channels · 256 Hz Nguyen2017_S: Imagery · 6 people · 64 channels · 256 Hz Nguyen2017_SL: Imagery · 6 people · 64 channels · 256 Hz Nguyen2017_V: Imagery · 8 people · 64 channels · 256 Hz — as MOABB lists it | “other-open” — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0108) | | [Nieto2022 ↗](https://openneuro.org/datasets/ds003626) — DOI: 10.1038/s41597-022-01147-2 · EEGDash: EEGDash/ds003626 | Imagery · 10 people · 136 channels · 1,024 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0109) | | [Ofner2017 ↗](https://bnci-horizon-2020.eu/database/data-sets) — DOI: 10.1371/journal.pone.0182578 | Imagery · 15 people · 96 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0110) | | [Ortiz2023 ↗](https://doi.org/10.6084/m9.figshare.21185362.v2) — DOI: 10.1038/s41597-023-02243-7 | Imagery · 10 people · 31 channels · 200 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0111) | | [Pan2023 ↗](https://doi.org/10.7910/DVN/251NOW) — DOI: 10.7910/DVN/251NOW | Imagery · 14 people · 28 channels · 250 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0112) | | [Pan2025 ↗](https://doi.org/10.7910/DVN/GH74ZG) — DOI: 10.7910/DVN/GH74ZG | Imagery · 10 people · 28 channels · 250 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0113) | | [PardoGarcia2026 ↗](https://doi.org/10.5281/zenodo.19599465) — DOI: 10.5281/zenodo.19599465 | Imagery · 18 people · 63 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0114) | | [Perdikis2018 ↗](https://doi.org/10.5281/zenodo.841764) — DOI: 10.5281/zenodo.841764 | Imagery · 2 people · 16 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0115) | | [PerezBlanco2026 ↗](https://doi.org/10.6084/m9.figshare.29666735) — DOI: 10.1038/s41597-026-07287-z | Imagery · 45 people · 16 channels · 512 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0116) | | [Peters2026 ↗](https://openneuro.org/datasets/ds007720) — DOI: 10.18112/openneuro.ds007720.v1.0.0 | P300 / ERP · 5 people · 23 channels · 300 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0117) | | [PoloHortiguela2025 ↗](https://doi.org/10.5281/zenodo.14672334) — DOI: 10.5281/zenodo.14672334 | Imagery · 6 people · 35 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0120) | | [Pressel2016 ↗](https://zenodo.org/records/19502780) — DOI: 10.1117/12.2255697 | Imagery · 15 people · 6 channels · 1,024 Hz — as MOABB lists it | “other-open” — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0121) | | [RomaniBF2025ERP ↗](https://zenodo.org/records/17225966) — DOI: 10.48550/arXiv.2510.10169 | P300 / ERP · 22 people · 8 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0123) | | [Rozado2015 ↗](https://doi.org/10.7910/DVN/28932) — DOI: 10.1371/journal.pone.0121262 | Imagery · 30 people · 32 channels · 512 Hz — as MOABB lists it | CC0 1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0124) | | [Schirrmeister2017 ↗](https://web.gin.g-node.org/robintibor/high-gamma-dataset/) — DOI: 10.1002/hbm.23730 | Imagery · 14 people · 128 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0125) | | [Schrag2026Pediatric ↗](https://doi.org/10.5281/zenodo.20848097) — DOI: 10.5281/zenodo.20848097 | SSVEP · 47 people · 16 channels · 256 Hz — as MOABB lists it | cc-by-4.0 — the deposit’s record | Step 1 of 5 · Listed — Status: Listed | Held: intended targets must be separated from the classifier outputs stored with them. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0126) | | [Shin2017 ↗](http://doc.ml.tu-berlin.de/hBCI) — MOABB: Shin2017A, Shin2017B · DOI: 10.1109/TNSRE.2016.2628057 | Shin2017A: Imagery · 29 people · 32 channels · 200 Hz Shin2017B: cognitive · 29 people · 32 channels · 200 Hz — as MOABB lists it | GPL-3.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0127) | | [Simoes2020 ↗](https://zenodo.org/records/19005186) — DOI: 10.3389/fnins.2020.568104 | P300 / ERP · 15 people · 8 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0129) | | [Sosulski2019 ↗](https://freidok.uni-freiburg.de/data/154576) — DOI: 10.48550/arXiv.2109.06011 | P300 / ERP · 13 people · 37 channels · 1,000 Hz — as MOABB lists it | CC-BY-SA-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0130) | | [Speier2017 ↗](https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/PHHHB6) — DOI: 10.1371/journal.pone.0175382 | P300 / ERP · 10 people · 32 channels · 256 Hz — as MOABB lists it | CC0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0131) | | [Tavakolan2017 ↗](https://zenodo.org/records/18967205) — DOI: 10.1371/journal.pone.0174161 | Imagery · 12 people · 32 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0133) | | [Thapa2025 ↗](https://doi.org/10.6084/m9.figshare.28632599) — DOI: 10.1038/s41597-025-06039-9 | Imagery · 23 people · 39 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0134) | | [Thielen2015 ↗](https://public.data.ru.nl/dcc/DSC_2018.00047_553_v3) — DOI: 10.1371/journal.pone.0133797 | c-VEP · 12 people · 64 channels · 2,048 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0135) | | [Thielen2021 ↗](https://doi.org/10.34973/9txv-z787) — DOI: 10.1088/1741-2552/abecef | c-VEP · 30 people · 8 channels · 512 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0136) | | [Vagaja2023 ↗](https://zenodo.org/records/8086086) — DOI: 10.5281/zenodo.8086086 | Imagery · 26 people · 38 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0138) | | [WRCC2023_MI_A ↗](https://doi.org/10.7910/DVN/J9JFES) — DOI: 10.7910/DVN/J9JFES | Imagery · 9 people · 59 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0139) | | [WRCC2023_MI_B ↗](https://doi.org/10.7910/DVN/HFNWRX) — DOI: 10.7910/DVN/HFNWRX | Imagery · 9 people · 59 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0140) | | [WRCC2023_MI_C ↗](https://doi.org/10.7910/DVN/G8FBHH) — DOI: 10.7910/DVN/G8FBHH | Imagery · 8 people · 59 channels · 1,000 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0141) | | [Wairagkar2018 ↗](https://researchdata.reading.ac.uk/117/) — DOI: 10.1371/journal.pone.0193722 | Imagery · 14 people · 19 channels · 1,024 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0142) | | [Wang2016 ↗](http://bci.med.tsinghua.edu.cn/download.html) — DOI: 10.1109/TNSRE.2016.2627556 | SSVEP · 34 people · 64 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0143) | | [Wang2021Combined ↗](https://zenodo.org/records/18873228) — DOI: 10.1111/ejn.15030 | SSVEP · 8 people · 33 channels · 1,000 Hz — as MOABB lists it | CC BY 4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0144) | | [Wang2025 ↗](https://huggingface.co/datasets?search=ZJU-MI-EEG) — DOI: 10.1109/TNSRE.2025.3591254 · Hugging Face: Jiaheng-Wang/ZJU-MI-EEG | Imagery · 15 people · 62 channels · 256 Hz — as MOABB lists it | ODC-BY — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0145) | | [Wang2026 ↗](https://kilthub.cmu.edu/articles/dataset/EEG_Datasets_for_Sensory-Guided_Joint_Learning_in_Motor_Imagery_Brain-Computer_Interfaces/32293995) — DOI: 10.1184/R1/32293995.v1 | Imagery · 39 people · 62 channels · 1,000 Hz — as MOABB lists it | CC-BY-NC-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0146) | | [Wang2026Handwriting ↗](https://doi.org/10.6084/m9.figshare.29987758.v4) — DOI: 10.1038/s41597-026-06708-3 | cognitive · 21 people · 32 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0147) | | [Weibo2014 ↗](http://dx.doi.org/10.7910/DVN/27306) — DOI: 10.1371/journal.pone.0114853 | Imagery · 10 people · 64 channels · 200 Hz — as MOABB lists it | CC0-1.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0148) | | [Wu2020 ↗](https://zenodo.org/records/18961128) — DOI: 10.1088/1741-2552/abc1b6 | Imagery · 6 people · 132 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0149) | | [Yagan2023 ↗](https://doi.org/10.17632/vyczny2r4w.1) — DOI: 10.1016/j.dsp.2023.103950 | P300 / ERP · 18 people · 32 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0150) | | [Yi2025 ↗](https://figshare.com/articles/dataset/Data/24123303) — DOI: 10.1038/s41597-025-05286-0 | Imagery · 18 people · 62 channels · 1,000 Hz — as MOABB lists it | CC-BY-NC-ND-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0152) | | [Zhang2017 ↗](https://doi.org/10.6084/m9.figshare.5579461.v1) — DOI: 10.1371/journal.pone.0188293 | Imagery · 12 people · 17 channels · 1,000 Hz — as MOABB lists it | CC BY 4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0153) | | [Zheng2020 ↗](https://figshare.com/articles/dataset/12824771) — DOI: 10.3389/fnins.2020.579469 | P300 / ERP · 14 people · 62 channels · 1,000 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0155) | | [Zhou2016 ↗](https://doi.org/10.6084/m9.figshare.2061654) — DOI: 10.1371/journal.pone.0162657 | Imagery · 4 people · 14 channels · 250 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0156) | | [Zhou2020 ↗](https://zenodo.org/records/18988317) — DOI: 10.3389/fnhum.2021.701091 | Imagery · 20 people · 28 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0157) | | [Zuo2025 ↗](https://figshare.com/articles/dataset/28740260) — DOI: 10.1038/s41597-025-05767-2 | Imagery · 30 people · 30 channels · 500 Hz — as MOABB lists it | CC-BY-4.0 — as MOABB lists it | Step 1 of 5 · Listed — Status: Listed | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-moabb-0158) | | **Corpora that require an agreement or request** · 6 rows — Corpora that require a signed agreement, a request or a registration. Their terms are quoted from a captured copy of the source’s own page. | | | | | | [AMIGOS ↗](https://www.eecs.qmul.ac.uk/mmv/datasets/amigos/download.html) | — | “The user may only use the dataset after this EULA has been signed and returned to the dataset administrators.” — [quoted from an archived copy of the source’s page, captured 19 October 2026 ↗](https://www.eecs.qmul.ac.uk/mmv/datasets/amigos/download.html) · SHA-256 `5e78e87238ab1c99` — Access: signed agreement | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-amigos) | | [DEAP ↗](https://www.eecs.qmul.ac.uk/mmv/datasets/deap/download.html) | — | “The user may only use the dataset after this EULA has been signed and returned to the dataset administrators.” — [quoted from an archived copy of the source’s page, captured 19 October 2026 ↗](https://www.eecs.qmul.ac.uk/mmv/datasets/deap/download.html) · SHA-256 `da59f1c39be9836b` — Access: signed agreement | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-deap) | | [DREAMER ↗](https://zenodo.org/records/546113) | — | “Please state your name, contact details (e-mail), institution, position, as well as the reason for requesting access to the DREAMER database.” — [quoted from the source’s page, captured 19 October 2026 ↗](https://zenodo.org/records/546113) · SHA-256 `1d1a718fa376c72c` — Access: on request | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-dreamer) | | [MAHNOB-HCI ↗](https://mahnob-db.eu/hci-tagging/media/doc/eula.pdf) — Also: MAHNOB-HCI-Tagging | — | “This document must be signed by a person with a permanent position at an academic institute (the signee).” — [quoted from the source’s page, captured 19 October 2026 ↗](https://mahnob-db.eu/hci-tagging/media/doc/eula.pdf) · SHA-256 `6bc468ee863eb01b` — Access: signed agreement | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-mahnob-hci) | | [SEED family ↗](https://bcmi.sjtu.edu.cn/home/seed/downloads.html) — Also: SEED, SEED-IV, SEED-V, SEED-VIG | — | “To gain the access to our datasets, please download and fill out the license agreement below and upload it on the APPLY page.” — [quoted from the source’s page, captured 19 October 2026 ↗](https://bcmi.sjtu.edu.cn/home/seed/downloads.html) · SHA-256 `ddc22ae373beb86e` — Access: signed agreement | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-seed) | | [TUH EEG Corpus ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/) — Also: TUEG, TUAB, TUSZ, TUEV, TUAR, TUEP, TUSL | — | “To request access to the TUH EEG Corpus, please fill out this form and email a signed copy” — [quoted from the source’s page, captured 19 October 2026 ↗](https://isip.piconepress.com/projects/nedc/html/tuh_eeg/) · SHA-256 `75fa8477215127f9` — Access: signed agreement | Step 1 of 5 · Listed — Status: Listed — Stays at step 1 under the access rule | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-corpus-tuh-eeg) | Download: [`open-register-20261106.json`](https://bci.report/data/open-register-20261106.json) (both ladders, with each step’s evidence and the findings as codes) and [`open-register-20261106.csv`](https://bci.report/data/open-register-20261106.csv) (this ladder, flat, with no findings). [Every file](https://bci.report/api/) A longer write-up of the Hugging Face rows is on the Hugging Face blog: [What we found listing the EEG datasets on Hugging Face ↗](https://huggingface.co/blog/Twu31/what-we-found-listing-the-eeg-datasets-on-hugging) ## Cite this page BCI Report (2026). *Every EEG dataset we know of, and how far we checked it*. https://bci.report/datasets/register/ Figures from release [`open-register-update-20261106`](https://bci.report/releases/#open-register-update-20261106) (2026-11-06). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/register/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Shin2022 · closed-loop motor imagery EEG (Figshare 20383716): EEG decoding results Closed-loop cursor control: the instructed right or left target, dry electrodes Shin2022 (Figshare 20383716) holds the recordings of a closed-loop motor imagery cursor study: 10 healthy participants, one recording day each, moved a cursor toward an instructed right or left target with visual feedback, recorded with a 16-channel dry-electrode g.tec g.Nautilus headset at 250 Hz. The release is organised into live and simulated recordings, each under three experimental conditions (BW, NT and CV), following the authors’ parallel design with a BCI simulator. Released by Hyonyoung Shin, Daniel Suma and Bin He under CC BY 4.0 and described in Frontiers in Human Neuroscience (2022). Only the live human recordings are used here. **Also known as** Shin2022 · Figshare 20383716 · Data from: Closed-loop motor imagery EEG simulation for brain-computer interfaces · Shin, Suma and He (2022) Description sources [doi.org](https://doi.org/10.6084/m9.figshare.20383716.v1) · [api.figshare.com](https://api.figshare.com/v2/articles/20383716/versions/1) · [www.frontiersin.org](https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2022.951591/full) ## Where it appears - [What do the files show once a public EEG dataset is checked end to end?](https://bci.report/topics/data-quality/#shin2022) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Data quality · the instructed right or left target, a participant held out entirely Source page: [Checked end to end](https://bci.report/topics/data-quality/#shin2022) · [data-quality-update.json](https://bci.report/data/data-quality-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Normalized spectral ridge | A participant held out entirely · the first second of feedback, signal-valid windows | Balanced accuracy | 52.59% (51.51%–53.66%) — Balanced accuracy pooled over all signal-valid windows: 52.55%. | 10 | | Retrospective duration control — comparator | A participant held out entirely · the whole feedback duration only, known after the trial ends (no EEG) | Balanced accuracy | 52.33% (50.87%–54.02%) — Balanced accuracy pooled over all signal-valid windows: 52.15%. | 10 | | Participant-weighted training prior — comparator | A participant held out entirely · no input | Balanced accuracy | 50.00% — Always the larger training class, each training participant weighted equally: every person gets the same score, so there is no interval. | 10 | Of 110 runs in the release’s human recordings, 9 pairs are byte-identical, leaving 101 distinct runs; 1 incomplete run is held whole. The 100 complete runs hold 2,400 closed feedback windows: 167 end within the first second; of the other 2,233, 13 are constant on at least one channel in their first second, leaving 2,220 signal-valid windows (92.5% of the closed windows), with both targets for all 10 participants. All three arms are scored on them, with participant-bootstrap intervals over 10,000 resamples. The scores cover the signal-valid windows only: feedback lasting at least one second, and a first second that is finite with positive 1–40 Hz power on all 16 channels. The label is the instructed target in closed-loop cursor recordings, which carry visual, controller and behaviour confounds. The duration control reads the whole feedback duration, known only after the trial ends, so it checks a confound; the spectral arm sits close to it, and no test between arms was run. Ten participants, one recording day each; the intervals are descriptive. ## Source and licence ### Shin2022 · closed-loop motor imagery EEG (Figshare 20383716) **Credit** Hyonyoung Shin, Daniel Suma and Bin He · Data from: Closed-loop motor imagery EEG simulation for brain-computer interfaces, figshare, doi:10.6084/m9.figshare.20383716.v1, CC BY 4.0. Paper: Hyonyoung Shin, Daniel Suma and Bin He, Closed-loop motor imagery EEG simulation for brain-computer interfaces, Frontiers in Human Neuroscience 16:951591 (2022), doi:10.3389/fnhum.2022.951591. Funding: NIH AT009263. Modifications: canonical deduplication of the byte-identical runs, first-second window selection and signal-validity filtering. Derived analysis by BCI Report; not endorsed by the authors. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.6084/m9.figshare.20383716.v1) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-shin2022) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Shin2022 · closed-loop motor imagery EEG (Figshare 20383716): EEG decoding results*. https://bci.report/datasets/shin2022/ Figures from release [`data-quality-update-20261020`](https://bci.report/releases/#data-quality-update-20261020) (2026-10-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/shin2022/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # TMNRED / ds005383: EEG decoding results Semantic target ERP TMNRED is a Chinese natural-reading EEG dataset from Tianjin University: 30 native Chinese-speaking participants read news headlines or short sentences containing target or non-target semantic items across 8 blocks, recorded with a portable 32-channel system (30 effective channels). Released on OpenNeuro (ds005383) and described by Bai et al. in Scientific Data (2025). **Also known as** TMNRED · ds005383 · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments · Bai et al. (2025) Description sources [openneuro.org](https://openneuro.org/datasets/ds005383) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC12032204/) · [doi.org](https://doi.org/10.1038/s41597-025-05036-2) ## Where it appears - [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json), [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) **Balanced accuracy, 10 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Spectral ridge: 51.0% (49.6%–52.5%) - Temporal ridge: 56.3% (54.4%–58.4%) - LaBraM: 53.2% (51.3%–54.9%) - CBraMod: 55.8% (54.2%–57.4%) - EEGNet: 61.4% (58.9%–64.0%) - XDAWNCov+TS+LR · 4 filters · 50 Hz: 60.3% (58.2%–62.6%) - xDAWN+LDA · 4 filters · 50 Hz: 56.8% (55.1%–58.6%) - ShallowFBCSPNet: 56.7% (54.7%–58.8%) - Deep4Net: 57.4% (55.4%–59.5%) - CSP+LDA: 52.7% (51.4%–54.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Reading · target versus nontarget events | Balanced accuracy | 51.0% (49.6%–52.5%) | 30 | | Temporal ridge | Reading · target versus nontarget events | Balanced accuracy | 56.3% (54.4%–58.4%) | 30 | | [LaBraM](https://bci.report/methods/labram/) | Reading · target versus nontarget events | Balanced accuracy | 53.2% (51.3%–54.9%) | 30 | | [CBraMod](https://bci.report/methods/cbramod/) | Reading · target versus nontarget events | Balanced accuracy | 55.8% (54.2%–57.4%) | 30 | | [EEGNet](https://bci.report/methods/eegnet/) | Reading · target versus nontarget events | Balanced accuracy | 61.4% (58.9%–64.0%) | 30 | | [XDAWNCov+TS+LR · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Reading · target versus nontarget events | Balanced accuracy | 60.3% (58.2%–62.6%) — Primary row · one fixed configuration (MOABB-reference settings) | 30 | | [xDAWN+LDA · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Reading · target versus nontarget events | Balanced accuracy | 56.8% (55.1%–58.6%) — Secondary row · one fixed configuration (MOABB-reference settings) | 30 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Reading · target versus nontarget events | Balanced accuracy | 56.7% (54.7%–58.8%) — One fixed seed and training budget, as the EEGNet cell | 30 | | [Deep4Net](https://bci.report/methods/deep4net/) | Reading · target versus nontarget events | Balanced accuracy | 57.4% (55.4%–59.5%) — Braindecode's automatically scaled Deep4Net for 1-s input (no temporal pooling, 229-sample classifier, 242,627 parameters), not the original Deep4Net configuration. | 30 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Reading · target versus nontarget events | Balanced accuracy | 52.7% (51.4%–54.0%) — Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. On this ERP protocol it reads band power, not the waveform. | 30 | ### Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) **Balanced accuracy, 15 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + ridge head: 53.7% (51.8%–55.4%) - REVE Large · Frozen encoder + ridge head: 59.7% (57.5%–61.8%) - LUNA Base · Frozen encoder + ridge head: 56.2% (54.2%–58.2%) - LUNA Large · Frozen encoder + ridge head: 57.1% (55.3%–58.9%) - CodeBrain · Frozen encoder + ridge head: 54.7% (53.0%–56.6%) - EEGMamba · Frozen encoder + ridge head: 54.4% (52.5%–56.3%) - ST-EEGFormer Base · Frozen encoder + ridge head: 54.6% (52.9%–56.4%) - ST-EEGFormer Large · Frozen encoder + ridge head: 57.8% (56.0%–59.6%) - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 54.8% (52.7%–56.9%) - ERP-FM Base · Frozen encoder + ridge head: 53.3% (51.7%–54.8%) - SingLEM · Frozen encoder + ridge head: 58.5% (56.1%–61.1%) - ZUNA 1.1 · Frozen encoder + ridge head: 53.9% (52.2%–55.6%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 55.8% (53.9%–57.6%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 54.9% (53.4%–56.4%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 55.0% (53.4%–56.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 53.7% (51.8%–55.4%) | 30 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 59.7% (57.5%–61.8%) | 30 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 56.2% (54.2%–58.2%) | 30 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 57.1% (55.3%–58.9%) | 30 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 54.7% (53.0%–56.6%) | 30 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 54.4% (52.5%–56.3%) | 30 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 54.6% (52.9%–56.4%) | 30 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 57.8% (56.0%–59.6%) | 30 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 54.8% (52.7%–56.9%) | 30 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 53.3% (51.7%–54.8%) | 30 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 58.5% (56.1%–61.1%) — Pretraining exposure: in the authors' published pretraining list. | 30 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 53.9% (52.2%–55.6%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 30 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 55.8% (53.9%–57.6%) | 30 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 54.9% (53.4%–56.4%) | 30 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 55.0% (53.4%–56.6%) | 30 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7); LUNA under CC BY-ND 4.0, with no modified weights shared; CodeBrain and ZUNA under Apache-2.0; EEGMamba, ST-EEGFormer and SingLEM under MIT; eeg-fm-masking under CC-BY-4.0; ERP-FM under CC BY-NC-SA 4.0, non-commercial. No model’s authors endorse these results. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) **BrainOmni Base: not run.** BrainTokenizer needs a 512-sample window (2.0 s at 256 Hz); the published segment is 1 s = 256 samples, so upstream unfold() would zero-pad 256 of 512 samples of every window (invented samples). All upstream downstream tasks use >= 2 s windows. ### EEGPT, frozen encoder · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 54.6% (53.3%–55.9%) | 30 | On the P300 and semantic targets, one-second windows ran as four real 0.25-second patches: the patch grid was set to four after the encoder was built, with no padding. EEGPT’s encoder treats each time patch on its own, so this is numerically the same as building it for 256-sample inputs, as the Orthosis pilot built it for its 128-sample windows. Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT: the checkpoint’s licence is unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. [Its licence and terms →](https://bci.report/protocols/semantic-target/#eegpt-terms) ## Source and licence ### TMNRED / ds005383 **Credit** Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0. [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-tmnred) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *TMNRED / ds005383: EEG decoding results*. https://bci.report/datasets/tmnred/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/tmnred/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Cattan PC/VR P300 dataset: EEG decoding results P300 on a PC screen and in a VR headset Visual P300 BCI EEG from 21 participants who used a P300-based BCI both on a personal computer and in a passive, smartphone-based virtual-reality headset, recorded with 16 electrodes at GIPSA-lab (Grenoble) in 2018 by Cattan, Andreev, Rodrigues and Congedo. Released on Zenodo (CC BY 4.0, 2019) and documented in arXiv:1903.11297; dataset ID VR.EEG.2018-GIPSA. **Also known as** Dataset of an EEG-based BCI experiment in Virtual Reality and on a Personal Computer · VR.EEG.2018-GIPSA · Cattan2019_VR · Zenodo 2605205 · arXiv:1903.11297 Description sources [zenodo.org](https://zenodo.org/records/2605205) · [arxiv.org](https://arxiv.org/abs/1903.11297) · [hal.science](https://hal.science/hal-02078533) · [moabb.neurotechx.com](https://moabb.neurotechx.com/docs/generated/moabb.datasets.Cattan2019_VR.html) ## Where it appears - [Does a P300 decoder calibrated on a screen still work in VR?](https://bci.report/topics/screen-to-vr/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Screen to VR · P300 across displays Source page: [Screen to VR](https://bci.report/topics/screen-to-vr/) · Chance level 50.0% · [context-update.json](https://bci.report/data/context-update.json) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Mean-window logistic regression · PC ⇄ VR, onset-corrected timing: 66.0% (61.6%–70.2%) - Spatiotemporal shrinkage LDA · PC ⇄ VR, onset-corrected timing: 63.6% (59.6%–67.7%) - Mean-window logistic regression · PC ⇄ VR, recorded-tag timing (sensitivity): 63.8% (61.0%–66.5%) - Spatiotemporal shrinkage LDA · PC ⇄ VR, recorded-tag timing (sensitivity): 57.2% (55.1%–59.4%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Mean-window logistic regression | PC ⇄ VR, onset-corrected timing | Balanced accuracy | 66.0% (61.6%–70.2%) | 21 | | Mean-window logistic regression | PC ⇄ VR, onset-corrected timing | AUROC | 0.732 (0.671–0.785) | 21 | | Spatiotemporal shrinkage LDA | PC ⇄ VR, onset-corrected timing | Balanced accuracy | 63.6% (59.6%–67.7%) | 21 | | Spatiotemporal shrinkage LDA | PC ⇄ VR, onset-corrected timing | AUROC | 0.713 (0.652–0.771) | 21 | | Mean-window logistic regression | PC ⇄ VR, recorded-tag timing (sensitivity) | Balanced accuracy | 63.8% (61.0%–66.5%) | 21 | | Mean-window logistic regression | PC ⇄ VR, recorded-tag timing (sensitivity) | AUROC | 0.719 (0.682–0.755) | 21 | | Spatiotemporal shrinkage LDA | PC ⇄ VR, recorded-tag timing (sensitivity) | Balanced accuracy | 57.2% (55.1%–59.4%) | 21 | | Spatiotemporal shrinkage LDA | PC ⇄ VR, recorded-tag timing (sensitivity) | AUROC | 0.641 (0.596–0.682) | 21 | ## Source and licence ### Cattan PC/VR P300 dataset **Credit** Grégoire Cattan, Anton Andreev, Pedro L. C. Rodrigues and Marco Congedo · Dataset of an EEG-based BCI experiment in Virtual Reality and on a Personal Computer, Zenodo, doi:10.5281/zenodo.2605205; documentation arXiv:1903.11297. [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.5281/zenodo.2605205) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-vr-pc-p300) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Cattan PC/VR P300 dataset: EEG decoding results*. https://bci.report/datasets/vr-pc-p300/ Figures from release [`context-update-20260927`](https://bci.report/releases/#context-update-20260927) (2026-09-27). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/vr-pc-p300/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # WBCIC-SHU motor imagery dataset: EEG decoding results Motor imagery across three recording sessions: two-class and three-class cohorts, kept apart The WBCIC-SHU motor imagery dataset (Figshare+ 22671172) holds EEG from 62 healthy participants, each recorded in three sessions with the same motor-imagery paradigm: 51 in a two-class task (left- or right-hand grasping) and 11 in a three-class task that adds foot-hooking. It was collected at the 2019 World Robot Conference BCI robot contest with a 64-channel wireless Neuracle cap at 1,000 Hz, and the release holds raw data, a processed derivative and code. Released by Banghua Yang and Fenqi Rong (Shanghai University) and described in Scientific Data (2025). Only the processed derivative of sessions 1 and 3 is used here. **Also known as** WBCIC-SHU · WBCIC-SHU Motor Imagery Dataset · WBCIC_SHU Motor Imagery dataset · Figshare+ 22671172 · A multi-day and high-quality EEG dataset for motor imagery brain-computer interface · Yang et al. (2025) Description sources [doi.org](https://doi.org/10.25452/figshare.plus.22671172.v5) · [doi.org](https://doi.org/10.1038/s41597-025-04826-y) · [api.figshare.com](https://api.figshare.com/v2/articles/22671172/versions/5) ## Where it appears - [Does a decoder trained on an earlier session still work later?](https://bci.report/topics/later-sessions/#wbcic-shu) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Later sessions · two-class motor imagery, session 1 to session 3 Source page: [Later sessions](https://bci.report/topics/later-sessions/#wbcic-shu) · Chance level 50.0% · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Relative spectral power + ridge: 53.9% (52.7%–55.1%) - CBraMod · frozen encoder, session-1 ridge readout: 67.7% (64.8%–70.5%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Source-majority prior — comparator | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 50.0% — The floor: one over the number of classes. | 51 | | Relative spectral power + ridge | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 53.9% (52.7%–55.1%) | 51 | | Relative spectral power + ridge | Minus the source prior, the same people and trials | Paired difference | +3.9 pp (+2.7 pp–+5.1 pp) | 51 | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 67.7% (64.8%–70.5%) | 51 | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Minus the relative spectral ridge, the same people and trials | Paired difference | +13.8 pp (+11.3 pp–+16.4 pp) — A comparison of two fixed pipelines on reused test trials; whether pretraining caused the gain, and whether this checkpoint saw WBCIC-SHU in pretraining, are open. | 51 | ### Later sessions · three-class motor imagery, session 1 to session 3 Source page: [Later sessions](https://bci.report/topics/later-sessions/#wbcic-shu) · Chance level 33.3% · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Relative spectral power + ridge: 37.6% (35.4%–40.0%) - CBraMod · frozen encoder, session-1 ridge readout: 51.6% (45.2%–58.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Source-majority prior — comparator | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 33.3% — The floor: one over the number of classes. | 11 | | Relative spectral power + ridge | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 37.6% (35.4%–40.0%) | 11 | | Relative spectral power + ridge | Minus the source prior, the same people and trials | Paired difference | +4.3 pp (+2.1 pp–+6.6 pp) | 11 | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 51.6% (45.2%–58.3%) | 11 | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Minus the relative spectral ridge, the same people and trials | Paired difference | +14.0 pp (+7.5 pp–+20.8 pp) — A comparison of two fixed pipelines on reused test trials; whether pretraining caused the gain, and whether this checkpoint saw WBCIC-SHU in pretraining, are open. | 11 | ## Source and licence ### WBCIC-SHU motor imagery dataset **Credit** Banghua Yang, Fenqi Rong, Yunlong Xie, Du Li, Jiayang Zhang, Fu Li, Guangming Shi and Xiaorong Gao · A multi-day and high-quality EEG dataset for motor imagery brain-computer interface, Scientific Data 12, 488 (2025), doi:10.1038/s41597-025-04826-y. Data: Banghua Yang and Fenqi Rong, WBCIC-SHU Motor Imagery Dataset, Figshare+, doi:10.25452/figshare.plus.22671172.v5, CC BY 4.0. Derived analysis by BCI Report; not endorsed by the authors. [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.25452/figshare.plus.22671172.v5) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-wbcic-shu) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *WBCIC-SHU motor imagery dataset: EEG decoding results*. https://bci.report/datasets/wbcic-shu/ Figures from release [`later-sessions-update-20261008`](https://bci.report/releases/#later-sessions-update-20261008) (2026-10-08). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/wbcic-shu/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Wearable SSVEP BCI dataset (dry and wet electrodes): EEG decoding results 12-target SSVEP, dry and wet electrodes An open wearable SSVEP dataset: 8-channel EEG from 102 healthy participants performing a cue-guided 12-target SSVEP task (9.25–14.75 Hz), recorded with wet and with dry electrodes, 10 consecutive blocks each. Participants were recruited through the BCI Brain-Controlled Robot Contest at the 2020 World Robot Contest. Released by Zhu et al. in Sensors (2021), with data on figshare. **Also known as** Wearable SSVEP BCI Dataset · An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces · figshare 13560281 · Zhu et al. (2021) Description sources [figshare.com](https://figshare.com/articles/dataset/_/13560281) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC7916479/) · [doi.org](https://doi.org/10.3390/s21041256) · [bci.med.tsinghua.edu.cn](https://bci.med.tsinghua.edu.cn/download.html) ## Where it appears - [Do dry EEG electrodes decode as well as wet ones?](https://bci.report/topics/dry-vs-wet/) - [How much calibration data does an EEG decoder need?](https://bci.report/topics/calibration-budget/) - [What lets a decoder answer an SSVEP frequency it never trained on?](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer) Follow-up tests that also run on these dry and wet headbands: [Pooling CCA’s per-harmonic statistics](https://bci.report/topics/new-ssvep-frequencies/#harmonic-pooling) · [Where the harmonic sum’s gain over joint CCA comes from](https://bci.report/topics/new-ssvep-frequencies/#harmonic-gap) · [One decay rule on harmonics and on filter-bank bands](https://bci.report/topics/new-ssvep-frequencies/#fixed-decay) · [A rising weight on the harmonics](https://bci.report/topics/new-ssvep-frequencies/#rising-weights) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### Dry vs. wet sensor transfer Source page: [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 16 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Uniform random · Wet recording · no source training: 8.6% (8.1%–9.1%) - Uniform random · Dry recording · no source training: 8.1% (7.6%–8.6%) - Same-frequency power · Wet recording · no source training: 39.8% (35.4%–44.3%) - Same-frequency power · Dry recording · no source training: 36.0% (31.8%–40.2%) - CCA · Wet recording · no source training: 81.6% (77.7%–85.1%) - CCA · Dry recording · no source training: 65.5% (60.6%–70.2%) - FBCCA · Wet recording · no source training: 82.3% (78.7%–85.7%) - FBCCA · Dry recording · no source training: 64.6% (59.5%–69.6%) - Spectral ridge · Trained on wet, tested on wet: 40.3% (36.5%–44.3%) - Spectral ridge · Trained on wet, tested on dry: 35.2% (31.8%–38.8%) - Spectral ridge · Trained on dry, tested on dry: 37.2% (33.5%–41.0%) - Spectral ridge · Trained on dry, tested on wet: 39.7% (35.9%–43.6%) - EEGNet · Trained on wet, tested on wet: 75.0% (70.3%–79.5%) - EEGNet · Trained on wet, tested on dry: 45.6% (40.5%–50.8%) - EEGNet · Trained on dry, tested on wet: 64.0% (59.1%–68.8%) - EEGNet · Trained on dry, tested on dry: 56.8% (51.9%–61.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Uniform random — comparator | Wet recording · no source training | Balanced accuracy | 8.6% (8.1%–9.1%) | 102 | | Uniform random — comparator | Dry recording · no source training | Balanced accuracy | 8.1% (7.6%–8.6%) | 102 | | Same-frequency power — comparator | Wet recording · no source training | Balanced accuracy | 39.8% (35.4%–44.3%) | 102 | | Same-frequency power — comparator | Dry recording · no source training | Balanced accuracy | 36.0% (31.8%–40.2%) | 102 | | [CCA](https://bci.report/methods/cca/) | Wet recording · no source training | Balanced accuracy | 81.6% (77.7%–85.1%) | 102 | | [CCA](https://bci.report/methods/cca/) | Dry recording · no source training | Balanced accuracy | 65.5% (60.6%–70.2%) | 102 | | [FBCCA](https://bci.report/methods/fbcca/) | Wet recording · no source training | Balanced accuracy | 82.3% (78.7%–85.7%) — Five Butterworth order-4 sub-bands (6–80 to 38–80 Hz), four harmonics | 102 | | [FBCCA](https://bci.report/methods/fbcca/) | Dry recording · no source training | Balanced accuracy | 64.6% (59.5%–69.6%) — Five Butterworth order-4 sub-bands (6–80 to 38–80 Hz), four harmonics | 102 | | Spectral ridge | Trained on wet, tested on wet | Balanced accuracy | 40.3% (36.5%–44.3%) | 102 | | Spectral ridge | Trained on wet, tested on dry | Balanced accuracy | 35.2% (31.8%–38.8%) | 102 | | Spectral ridge | Trained on dry, tested on dry | Balanced accuracy | 37.2% (33.5%–41.0%) | 102 | | Spectral ridge | Trained on dry, tested on wet | Balanced accuracy | 39.7% (35.9%–43.6%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on wet, tested on wet | Balanced accuracy | 75.0% (70.3%–79.5%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on wet, tested on dry | Balanced accuracy | 45.6% (40.5%–50.8%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on dry, tested on wet | Balanced accuracy | 64.0% (59.1%–68.8%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on dry, tested on dry | Balanced accuracy | 56.8% (51.9%–61.6%) | 102 | ### Calibration budget · spectral ridge Source page: [How much calibration?](https://bci.report/topics/calibration-budget/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 10 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Spectral ridge · Source-only, trained on wet, tested on wet: 39.2% (35.4%–43.0%) - Spectral ridge · Source-only, trained on wet, tested on dry: 35.4% (32.0%–39.0%) - Spectral ridge · Source-only, trained on dry, tested on wet: 39.0% (35.2%–42.9%) - Spectral ridge · Source-only, trained on dry, tested on dry: 37.2% (33.6%–40.9%) - Spectral ridge · 12 calibration trials · wet: 15.3% (13.7%–17.1%) - Spectral ridge · 24 calibration trials · wet: 18.2% (15.9%–20.7%) - Spectral ridge · 48 calibration trials · wet: 21.3% (18.3%–24.5%) - Spectral ridge · 12 calibration trials · dry: 15.0% (13.6%–16.5%) - Spectral ridge · 24 calibration trials · dry: 18.3% (16.3%–20.7%) - Spectral ridge · 48 calibration trials · dry: 21.8% (19.3%–24.5%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Spectral ridge | Source-only, trained on wet, tested on wet | Balanced accuracy | 39.2% (35.4%–43.0%) | 102 | | Spectral ridge | Source-only, trained on wet, tested on dry | Balanced accuracy | 35.4% (32.0%–39.0%) | 102 | | Spectral ridge | Source-only, trained on dry, tested on wet | Balanced accuracy | 39.0% (35.2%–42.9%) | 102 | | Spectral ridge | Source-only, trained on dry, tested on dry | Balanced accuracy | 37.2% (33.6%–40.9%) | 102 | | Spectral ridge | 12 calibration trials · wet | Balanced accuracy | 15.3% (13.7%–17.1%) | 102 | | Spectral ridge | 24 calibration trials · wet | Balanced accuracy | 18.2% (15.9%–20.7%) | 102 | | Spectral ridge | 48 calibration trials · wet | Balanced accuracy | 21.3% (18.3%–24.5%) | 102 | | Spectral ridge | 12 calibration trials · dry | Balanced accuracy | 15.0% (13.6%–16.5%) | 102 | | Spectral ridge | 24 calibration trials · dry | Balanced accuracy | 18.3% (16.3%–20.7%) | 102 | | Spectral ridge | 48 calibration trials · dry | Balanced accuracy | 21.8% (19.3%–24.5%) | 102 | ### Calibration budget · eTRCA and CCA Source page: [How much calibration?](https://bci.report/topics/calibration-budget/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 6 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CCA · No calibration · dry: 65.1% (60.0%–70.0%) - CCA · No calibration · wet: 81.1% (77.2%–84.8%) - Single-band eTRCA · 24 calibration trials · dry: 12.4% (10.4%–15.0%) - Single-band eTRCA · 24 calibration trials · wet: 39.8% (34.1%–45.8%) - Single-band eTRCA · 48 calibration trials · dry: 15.1% (12.6%–18.2%) - Single-band eTRCA · 48 calibration trials · wet: 64.0% (57.1%–70.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CCA](https://bci.report/methods/cca/) | No calibration · dry | Balanced accuracy | 65.1% (60.0%–70.0%) | 102 | | [CCA](https://bci.report/methods/cca/) | No calibration · wet | Balanced accuracy | 81.1% (77.2%–84.8%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 24 calibration trials · dry | Balanced accuracy | 12.4% (10.4%–15.0%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 24 calibration trials · wet | Balanced accuracy | 39.8% (34.1%–45.8%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 48 calibration trials · dry | Balanced accuracy | 15.1% (12.6%–18.2%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 48 calibration trials · wet | Balanced accuracy | 64.0% (57.1%–70.7%) | 102 | ### New SSVEP frequencies · CCA with a learned correction, trained on BETA and run unchanged Source page: [New SSVEP frequencies](https://bci.report/topics/new-ssvep-frequencies/#wearable-transfer) · Chance level 8.33% · [new-ssvep-frequencies-update.json](https://bci.report/data/new-ssvep-frequencies-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | CCA with learned correction | Dry electrodes · trained on BETA, run unchanged | Accuracy | 65.45% (60.82%–70.08%) | 102 | | Matched CCA — comparator | Dry electrodes · no training | Accuracy | 65.38% (60.73%–70.06%) | 102 | | CCA with learned correction | Dry electrodes · CCA with learned correction minus matched CCA, the same people and windows (primary contrast) | Paired difference, accuracy | +0.07 pp (−0.71–+0.91 pp) — The interval includes zero. | 102 | | Cropped-spectrum check (Wearable-102) — comparator | Dry electrodes · no training, a diagnostic only | Accuracy | 57.78% (52.79%–62.89%) | 102 | | CCA with learned correction | Wet electrodes · trained on BETA, run unchanged | Accuracy | 81.04% (77.14%–84.49%) | 102 | | Matched CCA — comparator | Wet electrodes · no training | Accuracy | 81.65% (77.60%–85.14%) | 102 | | CCA with learned correction | Wet electrodes · CCA with learned correction minus matched CCA, the same people and windows (primary contrast) | Paired difference, accuracy | −0.61 pp (−1.48–+0.23 pp) — The interval includes zero. | 102 | | Cropped-spectrum check (Wearable-102) — comparator | Wet electrodes · no training, a diagnostic only | Accuracy | 73.01% (68.46%–77.18%) | 102 | | CCA with learned correction | Both sensors · trained on BETA, run unchanged | Accuracy | 73.24% (69.40%–76.88%) | 102 | | Matched CCA — comparator | Both sensors · no training | Accuracy | 73.51% (69.71%–77.22%) | 102 | | CCA with learned correction | Both sensors · CCA with learned correction minus matched CCA, the same people and windows (secondary contrast) | Paired difference, accuracy | −0.27 pp (−0.95–+0.40 pp) — The interval includes zero. | 102 | | Cropped-spectrum check (Wearable-102) — comparator | Both sensors · no training, a diagnostic only | Accuracy | 65.40% (60.94%–69.69%) | 102 | The 105 models are CCA with a learned correction, trained on BETA (7 participant folds × 5 frequency rotations × 3 training seeds), run with no retraining on 102 people’s recordings as stored: 12 targets, 12,240 two-second windows per sensor. Each person’s accuracy is averaged over the 105 models, then over people, with no ensemble or model choice. Both methods are given the 12 candidate frequencies. The model’s score is a scaled CCA correlation plus a bounded learned term, so its difference from matched CCA is what the learned term adds. Intervals: paired person bootstrap, 2,000 draws, conditional on the fixed 105 models and unadjusted for the two sensors. Matched CCA uses sine and cosine references up to the third harmonic, as the models’ own CCA term does, and reached 65.38% dry and 81.65% wet; the dry-versus-wet page's CCA uses references up to the fourth harmonic on the same windows and reached 65.47% and 81.56%. Wearable-102’s stored amplitude unit is undocumented and the correction reads amplitude spectra, so a unit effect is not separated from the change of dataset. Window onset, hardware, reference, people and frequency grid all change at once; no equivalence margin was tested, and no one-factor ablation or target adaptation was run, so no single cause is isolated. Dry and wet are the primary contrasts; both sensors together and the cropped-spectrum check are secondary, the check a diagnostic only. This is a different task on other recordings, so these scores do not compare with BETA’s held-out or seen tasks. Wearable-102 also appears on other pages of this site, and its documentation cannot rule out that some people also appear in BETA. ## Source and licence ### Wearable SSVEP BCI dataset (dry and wet electrodes) **Credit** Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 CC BY 4.0 · [Dataset record ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) · [doi.org ↗](https://doi.org/10.3390/s21041256) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-wearable-ssvep-102) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Wearable SSVEP BCI dataset (dry and wet electrodes): EEG decoding results*. https://bci.report/datasets/wearable-ssvep-102/ Figures from releases [`new-ssvep-frequencies-update-20261020`](https://bci.report/releases/#new-ssvep-frequencies-update-20261020) (2026-10-20), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/wearable-ssvep-102/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # YSU asynchronous SSVEP-BCI dataset: EEG decoding results Asynchronous SSVEP, control and non-control states EEG dataset for studying asynchronous SSVEP-based brain-computer interfaces, with 24 participants, released by Zhao et al. (Yanshan University, Qinhuangdao) on figshare (24906300, CC BY 4.0; first posted 2023, revision 3 in 2024) and described in Brain-Apparatus Communication 3(1) (2024). **Also known as** YSU-an EEG dataset for studying asynchronous SSVEP-BCI · figshare 24906300 · An EEG dataset for studying asynchronous steady-state visual evoked potential (SSVEP) based brain computer interfaces Description sources [figshare.com](https://figshare.com/articles/dataset/_/24906300) · [doi.org](https://doi.org/10.1080/27706710.2024.2418650) ## Where it appears - [How often does an EEG decoder fire when nobody is giving a command?](https://bci.report/topics/when-not-to-act/) ## Published results Each figure comes from the reviewed download linked in its group. One row is above another only where their 95% intervals do not overlap. ### When not to act · 20 further people, two rejection rules Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [extension-update.json](https://bci.report/data/extension-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Fixed CCA with rejection | Global threshold, fixed on the pilot | Detection balanced accuracy | 76.0% (71.6%–80.2%) | 20 | | Fixed CCA with rejection | Global threshold, fixed on the pilot | Command windows accepted (coverage) | 748 / 960 | 20 | | Fixed CCA with rejection | Global threshold, fixed on the pilot | Accepted and correct | 654 / 960 | 20 | | Fixed CCA with rejection | Global threshold, fixed on the pilot | Correct among accepted, mean over people | 85.8% | 20 | | Fixed CCA with rejection | Global threshold, fixed on the pilot · no command · central image, flicker off | Accepted by mistake (per window) | 49 / 240 | 20 | | Fixed CCA with rejection | Global threshold, fixed on the pilot · no command · white wall, resting | Accepted by mistake (per window) | 93 / 240 | 20 | | Fixed CCA with rejection | Global threshold, fixed on the pilot · no command · central image, surround flickering | Accepted by mistake (per window) | 107 / 480 | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows | Detection balanced accuracy | 79.0% (74.2%–83.8%) | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows | Command windows accepted (coverage) | 746 / 960 | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows | Accepted and correct | 649 / 960 | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows | Correct among accepted, mean over people | 85.1% | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows · no command · central image, flicker off | Accepted by mistake (per window) | 32 / 240 | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows · no command · white wall, resting | Accepted by mistake (per window) | 79 / 240 | 20 | | Fixed CCA with rejection | Personal threshold, 96 own windows · no command · central image, surround flickering | Accepted by mistake (per window) | 79 / 480 | 20 | | Fixed CCA with rejection | Personal minus global, same people and windows | Paired difference, detection balanced accuracy | +3.0 pp (+0.5 pp–+5.7 pp) — 10 people improved, 8 got worse, 2 unchanged | 20 | ### When not to act · four-person development pilot Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [context-update.json](https://bci.report/data/context-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | Fixed CCA with rejection | Command intended | Frequency recognised | 152 / 192 | 4 | | Fixed CCA with rejection | Command intended | Accepted (coverage) | 130 / 192 | 4 | | Fixed CCA with rejection | Command intended | Accepted and correct | 121 / 192 | 4 | | Fixed CCA with rejection | No command · central image, flicker off | Accepted by mistake (per window) | 9 / 48 | 4 | | Fixed CCA with rejection | No command · looking at a white wall, resting | Accepted by mistake (per window) | 20 / 48 | 4 | | Fixed CCA with rejection | No command · central image while the surrounding stimuli flicker | Accepted by mistake (per window) | 11 / 96 | 4 | ## Source and licence ### YSU asynchronous SSVEP-BCI dataset **Credit** Jing Zhao, Qian Zhang, Xinrui Wang, Xueshuo Liu, Jiaxin Li, Fengjie Fan, Zhenhu Liang and Xiaoli Li · An EEG dataset for studying asynchronous steady-state visual evoked potential (SSVEP) based brain computer interfaces, Brain-Apparatus Communication 3(1) (2024), doi:10.1080/27706710.2024.2418650; data at doi:10.6084/m9.figshare.24906300.v3. [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) · [Dataset record ↗](https://doi.org/10.6084/m9.figshare.24906300.v3) BCI Report does not redistribute any recording. These are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. [Row in the open register →](https://bci.report/datasets/register/#r-site-ysu-async-ssvep) [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *YSU asynchronous SSVEP-BCI dataset: EEG decoding results*. https://bci.report/datasets/ysu-async-ssvep/ Figures from releases [`extension-update-20261002`](https://bci.report/releases/#extension-update-20261002) (2026-10-02), [`context-update-20260927`](https://bci.report/releases/#context-update-20260927) (2026-09-27). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/datasets/ysu-async-ssvep/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEG decoding methods with results on BCI Report Methods Every published figure for each method, by dataset. [The model register: every EEG model and paper we track, and how far we checked it (62 rows) →](https://bci.report/methods/register/) ## Methods with pages | Method | Family | Measured on | | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Compact model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 · Wearable SSVEP BCI dataset (dry and wet electrodes) · OpenBMI motor imagery (Lee et al. 2019) · BOAS · Bitbrain Open Access Sleep dataset | | [LaBraM](https://bci.report/methods/labram/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 · Mobile BCI dataset (SSVEP and ERP paradigms) | | [CBraMod](https://bci.report/methods/cbramod/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 · Mobile BCI dataset (SSVEP and ERP paradigms) · OpenBMI motor imagery (Lee et al. 2019) · BOAS · Bitbrain Open Access Sleep dataset · WBCIC-SHU motor imagery dataset | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Compact model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [Deep4Net](https://bci.report/methods/deep4net/) | Compact model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Classical method | ds003810 · EEGMAT · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [CCA](https://bci.report/methods/cca/) | Classical method | BETA · Wearable SSVEP BCI dataset (dry and wet electrodes) · Mobile BCI dataset (SSVEP and ERP paradigms) | | [FBCCA](https://bci.report/methods/fbcca/) | Classical method | BETA · Wearable SSVEP BCI dataset (dry and wet electrodes) | | [eTRCA](https://bci.report/methods/etrca/) | — | Wearable SSVEP BCI dataset (dry and wet electrodes) | | [TS+LR and MDM](https://bci.report/methods/riemannian/) | Classical method | ds003810 · EEGMAT · ds005342 | | [xDAWN+LDA and XDAWNCov+TS+LR](https://bci.report/methods/xdawn/) | Classical method | ds006593 · TMNRED / ds005383 | | [REVE](https://bci.report/methods/reve/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [LUNA](https://bci.report/methods/luna/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [BrainOmni](https://bci.report/methods/brainomni/) | Foundation model | ds003810 · EEGMAT · BETA · EESM19 scalp subset · ds005342 | | [CodeBrain](https://bci.report/methods/codebrain/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [ST-EEGFormer](https://bci.report/methods/st-eegformer/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [eeg-fm-masking](https://bci.report/methods/eeg-fm-masking/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [ERP-FM](https://bci.report/methods/erp-fm/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [SingLEM](https://bci.report/methods/singlem/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [ZUNA](https://bci.report/methods/zuna/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | | [EEGPT](https://bci.report/methods/eegpt/) | Foundation model | ds003810 · EEGMAT · BETA · ds006593 · TMNRED / ds005383 · EESM19 scalp subset · ds005342 | ## Other measured baselines BCI Report's own fixed baselines and references. Their figures sit in the topic pages and the core matrix. - Spectral ridge · Classical method - Temporal ridge · Classical method BCI Report’s own comparator ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. Beside the Spectral ridge row on seven core protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/#logpower-ridge), [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#logpower-ridge), [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#logpower-ridge), [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#logpower-ridge), [P300 target ERP](https://bci.report/protocols/p300-target/#logpower-ridge), [Semantic target ERP](https://bci.report/protocols/semantic-target/#logpower-ridge), [Sleep staging](https://bci.report/protocols/sleep-scalp/#logpower-ridge). ### How it differs from the Spectral ridge row | Aspect | Spectral ridge | Log band-power ridge | | --- | --- | --- | | Resolution | 90 bins of 0.5 Hz per channel (2-second and 30-second windows) or 45 bins of 1 Hz (1-second windows) | 5 broad bands per channel | | Normalisation | each channel’s 0.5–45 Hz power sums to 1 | one divisor per window, the median channel standard deviation, for all channels | | Clipping | none | ±15 median-channel standard deviations, before the spectrum | | Estimator | Welch with constant detrend: one Hann segment on 2-second and 1-second windows; 29 half-overlapping 2-second segments averaged on 30-second windows | one periodic-Hann periodogram over the whole window | | Window mean | removed per Welch segment | removed per window, before the input view | | Sampling rate | 250 Hz | 200 Hz | | Range | 0.5–45 Hz inclusive | 1–45 Hz, right-open | | Transform | log of the relative density per bin | log of the absolute band integral, in scaled units | | Head | the same scaler, alpha 100, solver, class weighting and folds | the same | On BETA the Spectral ridge row is SSVEP-specific: log1p of the Hann-windowed projection powers at the 40 target frequencies and 3 harmonics, 120 features per channel. This comparator keeps the same 5 broad bands, 5 features per channel, as a control stated before scoring. - Alpha is fixed at 100, as for every published ridge row; it was not tuned for 5 features per channel. Choosing the ridge alpha by the project's train-only rule (from 1, 100 and 10,000) in place of the fixed alpha 100 gave paired differences whose 95% intervals included zero on all seven protocols. - Computing the same 5 bands on the microvolt windows (250 Hz, no per-window scaling or clipping) in place of the scaled 200 Hz view gave paired differences whose 95% intervals included zero on six protocols; on the 4-channel BETA protocol the difference, scaled view minus microvolt view, was +0.39 pp (95% interval +0.06 to +0.74), a small difference. - On sleep staging the 30-second spectrum is one long periodogram; the Spectral ridge row averages 29 Welch segments of 2 seconds, and the frozen encoders average 15 segments of 2 seconds. - The project’s own code, run with numpy, scipy and scikit-learn (BSD-3-Clause); no third-party weights or code were run. Each protocol’s data keep their released record. - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) ## Scored beside the matrix or in a panel of their own Each links the section that holds its scores, with its comparator beside it. - Signal-JEPA · Foundation model · frozen, with a ridge head, on [motor imagery & rest](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) and [arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#signal-jepa-eegsimpleconv) → - EEGSimpleConv · Compact model · [trained from scratch on the same two protocols →](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) - NeuroLM-B · Foundation model · [frozen features on arithmetic & rest](https://bci.report/topics/pilot-studies/#neurolm-eegmat), and on [an infant case-control cohort →](https://bci.report/topics/clinical-groups/#iesseeg) - NeuroRVQ · Foundation model · [frozen backbone on three tasks →](https://bci.report/topics/pilot-studies/#neurorvq) - EEG-Captioner · Foundation model · [frozen encoder on ten-second arithmetic & rest windows →](https://bci.report/topics/pilot-studies/#eegmat-ten-second) - AXON · Foundation model · [frozen, on the same ten-second windows →](https://bci.report/topics/pilot-studies/#eegmat-ten-second) - LaBraM at Scale · Foundation model · [frozen encoders on every core protocol →](https://bci.report/topics/does-pretraining-help/#more-checkpoints) ## In the directory, not in the matrix Models listed in the model directory with no result in the core matrix or beside it, and the reason. - [BIOT ↗](https://github.com/ycq091044/BIOT) · Foundation model · Adapter needed - [BENDR ↗](https://huggingface.co/braindecode/braindecode-bendr) · Foundation model · Research candidate - [CSBrain ↗](https://github.com/yuchen2199/CSBrain) · Foundation model · Access unverified - [Neuro-GPT ↗](https://github.com/wenhui0206/NeuroGPT) · Foundation model · Research candidate - [EEG-Conformer ↗](https://github.com/eeyhsong/EEG-Conformer) · Compact model · Research candidate - MIRepNet · Foundation model · Catalogue only (not evaluated) - EEG-DINO · Foundation model · Catalogue only (not evaluated) [All datasets →](https://bci.report/datasets/) Catalogue ## Model directory: from compact baselines to foundation models Parameter counts depend on the backbone and task configuration. Foundation model Evaluated ### [LaBraM](https://bci.report/methods/labram/) 5.82M parameters A fixed configuration has been evaluated. See each task for channels and training mode. Status as released 2026-09-20 Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms [Official source ↗](https://github.com/935963004/LaBraM) Foundation model Evaluated ### [EEGPT](https://bci.report/methods/eegpt/) 25.29M parameters Evaluated locally, but no score is published in this release: the checkpoint license is unresolved. Results are withheld pending that review, not because the model failed to run. Released status: Rights review pending. Evaluated as a frozen encoder on the eight core protocols, in the same runs as the published LaBraM rows and, except on idle, the CBraMod rows, and in one pilot comparison outside every leaderboard. The checkpoint’s licence remains unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. Its core-matrix cells and pretraining-attribution rows are published on its page; the cells also sit beside the core matrix on every protocol page. The note above is the released one. [The pilot comparison →](https://bci.report/topics/pilot-studies/#orthosis) Status checked 2026-10-13 Code: Apache-2.0 · Checkpoint: license unresolved; excluded from new public scope (as released) [Official source ↗](https://github.com/BINE022/EEGPT) Foundation model Evaluated ### [CBraMod](https://bci.report/methods/cbramod/) 4.88M parameters A fixed configuration has been evaluated. See each task for channels and training mode. Status as released 2026-09-20 Code: MIT · Weights: Apache-2.0 (official model card) [Official source ↗](https://github.com/wjq-learning/CBraMod) Compact model Evaluated ### [EEGNet](https://bci.report/methods/eegnet/) 1.9K parameters A fixed configuration has been evaluated. See each task for channels and training mode. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) Compact model Evaluated ### [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) 30.5K parameters A fixed configuration has been evaluated. See each task for channels and training mode. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) Compact model Evaluated ### [Deep4Net](https://bci.report/methods/deep4net/) 274.6K parameters A fixed configuration has been evaluated. See each task for channels and training mode. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) Classical method Evaluated ### [CSP+LDA](https://bci.report/methods/csp-lda/) — parameters A fixed configuration has been evaluated. See each task for channels and training mode. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://mne.tools/stable/generated/mne.decoding.CSP.html) Foundation model Adapter needed ### BIOT 3.19M parameters Pretrained bipolar tokens do not directly match the unipolar recordings. A separately validated adapter is required. Status as released 2026-09-20 MIT [Official source ↗](https://github.com/ycq091044/BIOT) Foundation model Evaluated ### [REVE Base](https://bci.report/methods/reve/) 69.19M parameters Code and electrode positions are accessible. Base weights require accepting the official access agreement; they have not been downloaded. Released status: Access gated. The owner accepted the REVE Responsible Use License v1.0, and the foundation-model evaluation ran the base weights as a frozen probe on the eight core protocols and adapted them on EEGMAT. Aggregate scientific results only; no adapted weights are shared. The note above is the released one. [Source ↗](https://huggingface.co/brain-bzh/reve-base) Weights: REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Status checked 2026-10-04 MIT (code, as released) [Official source ↗](https://github.com/elouayas/reve_eeg) Foundation model Evaluated (two protocols, beside the matrix) ### Signal-JEPA 3.46M parameters Public, ungated weights; research candidate only; not downloaded or executed for this release. The official model card provides a 13,847,488-byte safetensors checkpoint trained on Lee2019 at 128 Hz. All 17 ds005342 channel names match its 62-channel pretraining layout, but a separate locked adaptation protocol is still required. Released status: Research candidate. Scored as a frozen encoder with a ridge head on motor imagery & rest and arithmetic & rest, beside the core matrix, with its own input adapter (0.5–40 Hz, 128 Hz). Aggregate results only; the weights are not redistributed. The note above is the released one. [Its rows on the two protocol pages →](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) Status checked 2026-11-06 MIT, as stated on the model card (as released) [Official source ↗](https://huggingface.co/braindecode/signal-jepa) Foundation model Research candidate ### BENDR 157.00M parameters Public, ungated weights; research candidate only; not downloaded or executed in this expansion. The Braindecode checkpoint is 628,580,476 bytes and expects 20 channels at 250 Hz. A documented montage or input adapter is required for 17-channel data. Status as released 2026-09-20 BSD-3-Clause on the Braindecode model card; verify original release terms before redistributing weights [Official source ↗](https://huggingface.co/braindecode/braindecode-bendr) Foundation model Access unverified ### CSBrain — parameters Public source and an official Google Drive weight link are listed; anonymous weight download, file size, and execution remain unverified. The backbone accepts caller-supplied brain regions and channel ordering, while the published BCIC head is fixed to 22 channels. A new adapter and a checkpoint-key loading audit are required. Status as released 2026-09-20 Unclear: no LICENSE file was found at the reviewed official repository commit, and no separate weight license was identified [Official source ↗](https://github.com/yuchen2199/CSBrain) Foundation model Research candidate ### Neuro-GPT — parameters Public code and Hugging Face weights; research candidate only; not downloaded or executed for this release. The model page shows an approximately 318 MB checkpoint. The published EEGConformer input path is fixed to 22 channels, so using 17 channels requires an explicit mapping or replacement of the spatial layer. Status as released 2026-09-20 GPL-3.0 for the official repository and model card [Official source ↗](https://github.com/wenhui0206/NeuroGPT) Compact model Research candidate ### EEG-Conformer — parameters Public source; no official pretrained checkpoint was found; not executed for this release. This is a from-scratch baseline rather than a foundation-model checkpoint. The author script fixes the spatial convolution to 22 channels and must be parameterized for other montages. Status as released 2026-09-20 GPL-3.0 [Official source ↗](https://github.com/eeyhsong/EEG-Conformer) Compact model Evaluated (two protocols, beside the matrix) ### EEGSimpleConv — parameters Public implementation; no pretrained weights; not executed for this release. The maintained Braindecode implementation accepts the channel count and sampling rate directly, making it a practical lightweight from-scratch comparator. Released status: Research candidate. Trained from scratch for 20 epochs on motor imagery & rest and arithmetic & rest, three seeds, beside the core matrix; the pages print the mean of the three seeds. The note above is the released one. [Its rows on the two protocol pages →](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) Status checked 2026-11-06 BSD-3-Clause for the Braindecode implementation; the separate author repository license is unclear (as released) [Official source ↗](https://braindecode.org/stable/generated/braindecode.models.EEGSimpleConv.html) Classical method Evaluated ### [Standard CCA](https://bci.report/methods/cca/) — parameters Fixed reference method; inspect each task for input features and fit protocol. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://scikit-learn.org/stable/) Classical method Evaluated ### Spectral ridge — parameters Fixed reference method; inspect each task for input features and fit protocol. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://scikit-learn.org/stable/) Classical method Evaluated ### Temporal ridge — parameters Fixed reference method; inspect each task for input features and fit protocol. Status as released 2026-09-20 Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. [Official source ↗](https://scikit-learn.org/stable/) Foundation model Evaluated ### [REVE Large](https://bci.report/methods/reve/) 389.98M parameters Frozen probes on all eight core protocols and the EEGMAT adaptation, with the same corpus and terms as REVE Base. Weights: REVE Responsible Use License v1.0 (as REVE Base). Status checked 2026-10-04 REVE Responsible Use License v1.0 (owner-approved 2026-10-04) [Official source ↗](https://arxiv.org/abs/2510.21585) Foundation model Evaluated ### [LUNA (Base, Large)](https://bci.report/methods/luna/) 6.57M · 40.45M parameters Base · Large Frozen probes on all eight core protocols; Base was also adapted on EEGMAT, Large was not, by design. Weights: CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. Status checked 2026-10-04 CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) [Official source ↗](https://arxiv.org/abs/2510.22257) Foundation model Evaluated (6 of 8 protocols) ### [BrainOmni Base](https://bci.report/methods/brainomni/) 31.84M parameters Frozen probes on six of the eight core protocols and the EEGMAT adaptation. The two one-second ERP protocols were not run: its tokenizer needs two-second windows. Status checked 2026-10-04 MIT [Official source ↗](https://arxiv.org/abs/2505.18185) Foundation model Evaluated ### [CodeBrain](https://bci.report/methods/codebrain/) 13.53M parameters Frozen probes on all eight core protocols and the EEGMAT adaptation, pooled as the published CBraMod rows are. Status checked 2026-10-04 Apache-2.0 [Official source ↗](https://arxiv.org/abs/2506.09110) Foundation model Evaluated ### [EEGMamba](https://bci.report/methods/eegmamba/) 3.28M parameters The Neural Networks (2025) model from the CBraMod group. Frozen probes on all eight core protocols and the EEGMAT adaptation. Status checked 2026-10-04 MIT [Official source ↗](https://doi.org/10.1016/j.neunet.2025.107816) Foundation model Evaluated ### [ST-EEGFormer (Base, Large)](https://bci.report/methods/st-eegformer/) 85.18M · 303.00M parameters Base · Large Frozen probes on all eight core protocols; Base was also adapted on EEGMAT. BETA is in the authors’ published pretraining list, so both BETA protocols are flagged. Weights: MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). Status checked 2026-10-04 MIT (repo LICENSE; README reserves the paper, diagrams and the name) [Official source ↗](https://openreview.net/forum?id=5Xwm8e6vbh) Foundation model Evaluated (frozen probes, masking ablation) ### [eeg-fm-masking (4 checkpoints)](https://bci.report/methods/eeg-fm-masking/) 12.69M parameters Four checkpoints of one masking study, as frozen probes on all eight core protocols: the paper-recommended MAE checkpoint as a matrix row, three siblings in a masking ablation. Weights: CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). Status checked 2026-10-04 CC-BY-4.0 (cite arXiv:2609.33487) [Official source ↗](https://arxiv.org/abs/2609.33487) Foundation model Evaluated (frozen probes) ### [ERP-FM Base](https://bci.report/methods/erp-fm/) 2.02M parameters Frozen probes on all eight core protocols: the two ERP protocols are in its design, the other six are negative controls. Weights: CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. Status checked 2026-10-04 CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) [Official source ↗](https://arxiv.org/abs/2609.32796) Foundation model Evaluated ### [SingLEM](https://bci.report/methods/singlem/) 3.27M parameters A single-channel encoder. Frozen probes on all eight core protocols and the EEGMAT adaptation. TMNRED is in the authors’ published pretraining list, so the semantic-target protocol is flagged. Weights: MIT. Status checked 2026-10-04 MIT (checkpoint committed in the MIT repository) [Official source ↗](https://arxiv.org/abs/2509.17920) Foundation model Evaluated ### [ZUNA 1.1](https://bci.report/methods/zuna/) 172.10M parameters Frozen probes on all eight core protocols and the EEGMAT adaptation. Its authors publish no pretraining list, so every cell is marked exposure unknown. Weights: Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. Status checked 2026-10-04 Apache-2.0; model card: research use only, not for diagnosis or clinical use [Official source ↗](https://arxiv.org/abs/2607.27308) Foundation model Catalogue only (not evaluated) ### MIRepNet — parameters Listed only, by the owner’s decision. Status checked 2026-10-04 No licence reviewed Foundation model Catalogue only (not evaluated) ### EEG-DINO — parameters Listed only, by the owner’s decision. The upstream weights carry no licence. Status checked 2026-10-04 No licence reviewed External ## Field notes: other people’s work Curated research updates, linked to original sources. Benchmark 2026-08-03 ### [A new benchmark compares EEG foundation models with task-specific decoders ↗](https://arxiv.org/abs/2601.17883v3) Researchers surveyed 55 representative models and evaluated 12 open-source foundation models with task-specific baselines across 13 datasets and nine BCI paradigms. The study reports that linear probing is often insufficient and that larger models do not automatically generalize better, reinforcing the need for fixed cross-subject and few-shot calibration protocols. Original arXiv paper Model release 2025-10-24 ### [REVE releases an EEG foundation model for varying electrode layouts ↗](https://arxiv.org/abs/2510.21585) REVE reports pretraining on 92 datasets, about 60,000 hours of EEG, and 25,000 participants, with position encodings designed for different electrode arrangements. Its public code, weights, and tutorials make cross-montage transfer an auditable evaluation target. Access to the Base checkpoint is gated; it is not yet available in our local test pool. REVE Base and REVE Large have since been evaluated here; their frozen-probe rows are on every protocol page. [REVE’s method page →](https://bci.report/methods/reve/) Original arXiv paper and NeurIPS 2025 paper page Research release 2025-09-18 ### [NeurIPT brings heterogeneous EEG configurations into a unified pretraining framework ↗](https://openreview.net/forum?id=D4hcJPkJ3y) NeurIPT models homogeneous and heterogeneous spatiotemporal relationships across BCI tasks and electrode layouts and appeared as a NeurIPS 2025 poster. Its official repository still says that code restoration is in progress, so it is a research update rather than a reproducible model card. OpenReview / NeurIPS 2025 Model release 2025-06-29 ### [CSBrain adds cross-scale spatiotemporal structure to general EEG decoding ↗](https://arxiv.org/abs/2506.23075) CSBrain combines local temporal windows, brain regions, and structured sparse attention, with reported experiments spanning 16 datasets and 11 task types. The official repository provides pretraining and downstream scripts plus a public weight entry point, making it a candidate for a later unified-protocol evaluation. Original arXiv paper and NeurIPS 2025 paper page Real-time BCI 2025-06-02 ### [An eight-channel, low-preprocessing self-supervised EEG model targets real-time BCI ↗](https://arxiv.org/abs/2506.01867) This study applies HuBERT-style self-supervised learning to eight-channel scalp EEG and examines motor imagery, P300, subject differences, and alpha rhythms. It motivates a separate deployment-oriented track for low channel counts, limited preprocessing, and runtime budgets. Original arXiv paper --- Markdown copy of https://bci.report/methods/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # BrainOmni: results on public EEG datasets Foundation model BrainOmni is a pretrained foundation model for EEG and MEG, released in tiny and base sizes, with a BrainTokenizer front end; its paper's appendix lists the EEG and MEG recordings it was pretrained on. BCI Report ran Base frozen on six of the eight core protocols: its tokenizer needs two-second windows, so the two one-second ERP protocols were not run. It keeps its upstream common average reference on the small layouts here. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** BrainOmni Base · BrainTokenizer · OpenTSLab/BrainOmni **Reference** [BrainOmni, arXiv:2505.18185 ↗](https://arxiv.org/abs/2505.18185) Directory status: Evaluated (6 of 8 protocols) · Description sources [arxiv.org](https://arxiv.org/abs/2505.18185v3) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 58.8% (57.0%–60.8%) | 10 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 61.6% (58.4%–64.9%) | 36 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - BrainOmni Base · Frozen encoder + trained head: 64.5% (59.6%–69.2%) - BrainOmni Base · LoRA rank 4 + head: 64.8% (59.6%–69.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + trained head | Balanced accuracy | 64.5% (59.6%–69.2%) — 16,386 trainable parameters. | 36 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | LoRA rank 4 + head | Balanced accuracy | 64.8% (59.6%–69.7%) — 106,498 trainable parameters. | 36 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | LoRA minus frozen + head, same people and folds | Paired difference | +0.3 pp (−1.4 pp–+1.8 pp) — 18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established. | 36 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 21.7% (18.6%–25.0%) | 70 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 13.0% (10.6%–15.6%) | 70 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% **BrainOmni Base: not run.** BrainTokenizer needs a 512-sample window (2.0 s at 256 Hz); the published segment is 1 s = 256 samples, so upstream unfold() would zero-pad 256 of 512 samples of every window (invented samples). All upstream downstream tasks use >= 2 s windows. ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% **BrainOmni Base: not run.** BrainTokenizer needs a 512-sample window (2.0 s at 256 Hz); the published segment is 1 s = 256 samples, so upstream unfold() would zero-pad 256 of 512 samples of every window (invented samples). All upstream downstream tasks use >= 2 s windows. ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + ridge head | Balanced accuracy | 67.7% (65.8%–69.5%) | 20 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + linear head | Commands detected ≤3 s | 13 / 60 — 1 of 4 people always abstained. | 4 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | Frozen encoder + linear head | Idle false activations | 3 / 60 | 4 | Weights terms: BrainOmni under MIT. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### BrainOmni Base - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 31.84M - Revision: `OpenTSLab/BrainOmni @ 9a4d3c70` - Checkpoint SHA-256: `435db24e57a55df05aa7e16355def7b7ecbedb22aa1ec16063e7d14efd2386d0` - Paper: [paper ↗](https://arxiv.org/abs/2505.18185) - Weights licence: MIT - Rights review: Permissive weights licence; aggregate scores only, no weights redistributed. - Row footnote: p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2505.18185v3) | ## Cite this page BCI Report (2026). *BrainOmni: results on public EEG datasets*. https://bci.report/methods/brainomni/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/brainomni/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # CBraMod: results on public EEG datasets Foundation model CBraMod is an EEG foundation model from Wang et al. (Zhejiang University; ICLR 2025). Its criss-cross Transformer models spatial and temporal dependencies between EEG patches with two parallel attention branches, uses an asymmetric conditional positional encoding, and is pretrained by masked patch reconstruction on the TUEG clinical corpus (over 9,000 hours retained after preprocessing). Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/); [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/); [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/); [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/); [WBCIC-SHU motor imagery dataset](https://bci.report/datasets/wbcic-shu/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** CBraMod · Criss-Cross Brain Foundation Model · weighting666/CBraMod · Wang et al. 2025 **Reference** [Wang J, Zhao S, Luo Z, Zhou Y, Jiang H, Li S, Li T, Pan G. CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding. International Conference on Learning Representations (ICLR), 2025. ↗](https://arxiv.org/abs/2412.07236) Directory status: Evaluated · [Official code ↗](https://github.com/wjq-learning/CBraMod) · Description sources [arxiv.org](https://arxiv.org/abs/2412.07236) · [github.com](https://github.com/wjq-learning/CBraMod) · [huggingface.co](https://huggingface.co/weighting666/CBraMod) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Rest versus right-hand imagery | Balanced accuracy | 62.9% (59.6%–66.5%) | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · pretrained | Balanced accuracy | 62.9% | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · random initialization (mean) | Balanced accuracy | 56.5% | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · pretrained | Balanced accuracy | 62.7% | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | MI / rest · random initialization (mean) | Balanced accuracy | 54.9% | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Reliable decisions · when to decline a decision (crude: ten people) Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 7 / 1,200 — 6 of 10 people had nothing accepted. | 10 | | [CBraMod](https://bci.report/methods/cbramod/) | Learned reject option minus calibrated confidence, same classifier (a secondary contrast) | Error among the 80% most certain, difference | +0.7 pp (−0.8 pp–+2.4 pp) — No difference resolved: the interval includes zero. | 10 | **Limits:** Ten people, two per test fold. Bootstrap intervals are crude; no person-level dispersion is published. Balanced, so no deployment rate. Its learned-reject contrasts are secondary. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Serial subtraction versus resting EEG | Balanced accuracy | 62.3% (58.1%–66.5%) | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · pretrained | Balanced accuracy | 62.3% | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · random initialization (mean) | Balanced accuracy | 56.4% | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · pretrained | Balanced accuracy | 68.4% | 36 | | [CBraMod](https://bci.report/methods/cbramod/) | Mental workload · random initialization (mean) | Balanced accuracy | 61.7% | 36 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 33.7% (29.8%–37.6%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 27.6% (24.1%–31.3%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Fixed threshold: calibrated confidence at least 0.8 | Trials accepted | 165 / 11,200 — 38 of 70 people had nothing accepted. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | −0.03 pp (−0.8 pp–+0.6 pp) — No difference resolved: the interval includes zero. | 70 | **Limits:** Uniform target prior by design; cue-paced lab spelling with no idle state, so no false-activation figure of any kind. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [BETA](https://bci.report/datasets/beta/) · Questions in language · SSVEP (route 3) Source page: [Questions in language](https://bci.report/topics/questions-in-language/) · [questions-in-language-update.json](https://bci.report/data/questions-in-language-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · label template − question number (TPL − ID) · seen, 32-way · three seeds | Paired difference, language arm minus reference (pp) | −3.49 pp (−4.10–−2.93 pp) — ID higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · description − question number (DESC − ID) · seen, 32-way · three seeds | Paired difference, language arm minus reference (pp) | −3.37 pp (−3.98–−2.80 pp) — ID higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new sentence frame − training wording (TPL(a1) − TPL(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −2.72 pp (−4.43–−1.43 pp) — TPL(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · synonym − training wording (TPL(a2) − TPL(train)) · seen, reworded, for these three synonyms · three seeds | Paired difference, language arm minus reference (pp) | −1.33 pp (−1.53–−1.14 pp) — TPL(train) higher; equivalent within ±2 pp; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new description − training descriptions (DESC(a3) − DESC(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −2.95 pp (−3.81–−2.26 pp) — DESC(train) higher; 2 pp margin not met; passes the triviality canary. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − CCA, no training (TPL − CCA) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −61.94 pp (−65.65–−58.06 pp) — Unseen question. CCA higher; 2 pp margin not met; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − average of the two neighbouring seen frequencies (TPL − NN(TPL)) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −1.16 pp (−1.97–−0.37 pp) — Unseen question. NN(TPL) higher; difference only: no margin; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − numeric code (TPL − NUM) · unseen frequency, 8-way · three seeds | Paired difference, language arm minus reference (pp) | −2.46 pp (−3.23–−1.69 pp) — Unseen question. NUM higher; difference only: no margin; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − CCA, no training (TPL − CCA) · unseen frequency, neighbour 3-way (frequency and stimulus phase) · three seeds | Paired difference, language arm minus reference (pp) | −54.67 pp (−57.77–−51.47 pp) — Unseen question. CCA higher; 2 pp margin not met; headroom gate passed. | 70 | | [CBraMod](https://bci.report/methods/cbramod/) | P4 · label template − numeric code (TPL − NUM) · unseen frequency, neighbour 3-way (frequency and stimulus phase) · three seeds | Paired difference, language arm minus reference (pp) | +15.61 pp (+14.25–+16.93 pp) — Unseen question. TPL higher; difference only: no margin; headroom gate passed. | 70 | **Limits:** 'Unseen' means unseen by the EEG head; the text encoder has read these words, and every token of a held-out frequency also occurs in the training texts. Held-out SSVEP questions name a frequency, which CCA answers without training; interpolation tested, extrapolation secondary. In the neighbour 3-way the unseen frequency competes against trained detectors for its neighbours; at L1 that test also involves stimulus phase. Rewording and unseen results are separate rows and never pooled. 35 pre-declared comparisons, not corrected for multiplicity. The pre-run permutation canary failed twice (revisions 3 and 4) and was revised by the owner, who overrode revision 4's "no further revision"; revision 5 passed. Both failed records are kept and disclosed. [Every limitation of route 3 →](https://bci.report/topics/questions-in-language/#methods-and-limits) ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Visual target versus nontarget events | Balanced accuracy | 55.0% (53.3%–56.7%) | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Reading · target versus nontarget events | Balanced accuracy | 55.8% (54.2%–57.4%) | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 71.3% (68.7%–73.7%) | 20 | ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 18.3% | 4 | | [CBraMod](https://bci.report/methods/cbramod/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% | 4 | ### [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) · Movement · SSVEP, 2-second windows Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 3 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CBraMod · Standing · scalp · 8 ch: 59.9% (54.3%–65.4%) - CBraMod · Slow walk · 0.8 m/s · scalp · 8 ch: 46.2% (41.1%–51.4%) - CBraMod · Fast walk · 1.6 m/s · scalp · 8 ch: 46.0% (42.2%–49.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | Standing · scalp · 8 ch | Balanced accuracy | 59.9% (54.3%–65.4%) | 23 | | [CBraMod](https://bci.report/methods/cbramod/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 46.2% (41.1%–51.4%) | 23 | | [CBraMod](https://bci.report/methods/cbramod/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 46.0% (42.2%–49.6%) | 23 | ### [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/) · One model, several questions · motor imagery (route 2) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · Chance level 50.0% · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) **Balanced accuracy, 6 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CBraMod · fixed heads (B-lin-fz-sgd) · imagery or rest (MI-A) · heads on frozen features, three seeds: 69.2% (67.0%–71.5%) - CBraMod · shared hidden layer (B-sh-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds: 70.9% (68.4%–73.4%) - CBraMod · question-conditioned head (C1-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds: 70.8% (68.4%–73.2%) - CBraMod · fixed heads (B-lin-fz-sgd) · which hand (MI-B) · heads on frozen features, three seeds: 56.1% (54.8%–57.4%) - CBraMod · shared hidden layer (B-sh-fz) · which hand (MI-B) · heads on frozen features, three seeds: 56.7% (55.4%–58.2%) - CBraMod · question-conditioned head (C1-fz) · which hand (MI-B) · heads on frozen features, three seeds: 57.2% (55.8%–58.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · imagery or rest (MI-A) · heads on frozen features, three seeds | Balanced accuracy | 69.2% (67.0%–71.5%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds | Balanced accuracy | 70.9% (68.4%–73.4%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · imagery or rest (MI-A) · heads on frozen features, three seeds | Balanced accuracy | 70.8% (68.4%–73.2%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | imagery or rest (MI-A) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +1.54 pp (+0.86–+2.30 pp) — C1-fz higher; 2 pp margin not met; gate passed. | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · which hand (MI-B) · heads on frozen features, three seeds | Balanced accuracy | 56.1% (54.8%–57.4%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · which hand (MI-B) · heads on frozen features, three seeds | Balanced accuracy | 56.7% (55.4%–58.2%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · which hand (MI-B) · heads on frozen features, three seeds | Balanced accuracy | 57.2% (55.8%–58.6%) | 51 | | [CBraMod](https://bci.report/methods/cbramod/) | which hand (MI-B) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +1.09 pp (+0.08–+2.12 pp) — C1-fz higher; 2 pp margin not met (not read: at floor); at floor: reported, not counted. | 51 | **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. The OpenBMI numbers here use route 2's own 2-s windows, label-free high-pass and person-disjoint folds; they are not comparable with the site's OpenBMI cross-session calibration figures. A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ### [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) · One model, several questions · sleep (route 2) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) **Balanced accuracy, 9 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CBraMod · fixed heads (B-lin-fz-sgd) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds: 70.7% (68.5%–72.7%) - CBraMod · shared hidden layer (B-sh-fz) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds: 71.4% (69.1%–73.6%) - CBraMod · question-conditioned head (C1-fz) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds: 70.9% (68.6%–73.0%) - CBraMod · fixed heads (B-lin-fz-sgd) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds: 63.5% (62.3%–64.7%) - CBraMod · shared hidden layer (B-sh-fz) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds: 65.5% (64.1%–66.7%) - CBraMod · question-conditioned head (C1-fz) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds: 65.2% (63.9%–66.5%) - CBraMod · fixed heads (B-lin-fz-sgd) · first or second half of the night (SL-F) · heads on frozen features, three seeds: 64.5% (62.8%–66.1%) - CBraMod · shared hidden layer (B-sh-fz) · first or second half of the night (SL-F) · heads on frozen features, three seeds: 65.2% (63.6%–66.9%) - CBraMod · question-conditioned head (C1-fz) · first or second half of the night (SL-F) · heads on frozen features, three seeds: 65.2% (63.6%–66.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds | Balanced accuracy | 70.7% (68.5%–72.7%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds | Balanced accuracy | 71.4% (69.1%–73.6%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · sleep stage, five stages (SL-A) · heads on frozen features, three seeds | Balanced accuracy | 70.9% (68.6%–73.0%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | sleep stage, five stages (SL-A) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +0.19 pp (−0.74–+1.14 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds | Balanced accuracy | 63.5% (62.3%–64.7%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds | Balanced accuracy | 65.5% (64.1%–66.7%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · next epoch’s stage differs (SL-E) · heads on frozen features, three seeds | Balanced accuracy | 65.2% (63.9%–66.5%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | next epoch’s stage differs (SL-E) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +1.75 pp (+0.92–+2.63 pp) — C1-fz higher; 2 pp margin not met; gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | fixed heads (B-lin-fz-sgd) · first or second half of the night (SL-F) · heads on frozen features, three seeds | Balanced accuracy | 64.5% (62.8%–66.1%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | shared hidden layer (B-sh-fz) · first or second half of the night (SL-F) · heads on frozen features, three seeds | Balanced accuracy | 65.2% (63.6%–66.9%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | question-conditioned head (C1-fz) · first or second half of the night (SL-F) · heads on frozen features, three seeds | Balanced accuracy | 65.2% (63.6%–66.9%) | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | first or second half of the night (SL-F) · conditioned head minus fixed heads, on frozen CBraMod features (C1-fz − B-lin-fz-sgd) | Paired difference, balanced accuracy | +0.77 pp (−0.14–+1.67 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 100 | **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. SL-E and SL-F are largely predictable from the current stage: read off the true current stage alone, they reach 61.9% and 59.2% balanced accuracy (100 people; descriptive). A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ### [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) · Questions in language · sleep (route 3) Source page: [Questions in language](https://bci.report/topics/questions-in-language/) · [questions-in-language-update.json](https://bci.report/data/questions-in-language-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · label template − question number (TPL − ID) · seen, 5-way · three seeds | Paired difference, language arm minus reference (pp) | −0.05 pp (−0.70–+0.61 pp) — No difference shown; equivalent within ±2 pp; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P1 · description − question number (DESC − ID) · seen, 5-way · three seeds | Paired difference, language arm minus reference (pp) | +0.02 pp (−0.59–+0.62 pp) — No difference shown; equivalent within ±2 pp; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new sentence frame − training wording (TPL(a1) − TPL(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −0.46 pp (−1.33–+0.16 pp) — No difference shown; equivalent within ±2 pp; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · synonym − training wording (TPL(a2) − TPL(train)) · seen, reworded, for these three synonyms · three seeds | Paired difference, language arm minus reference (pp) | −12.60 pp (−13.74–−11.43 pp) — TPL(train) higher; 2 pp margin not met; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P2 · new description − training descriptions (DESC(a3) − DESC(train)) · seen, reworded · three seeds | Paired difference, language arm minus reference (pp) | −20.87 pp (−29.49–−13.71 pp) — DESC(train) higher; 2 pp margin not met; passes the triviality canary. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · asleep · implicit name · three seeds | log R, the head’s remaining error over the read-off’s | +3.520 (+3.139–+3.967) — Unseen question. R 33.80; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · asleep · explicit name · three seeds | log R, the head’s remaining error over the read-off’s | +1.951 (+1.623–+2.357) — Unseen question. R 7.04; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · N1, N2 or N3 sleep · explicit name · three seeds | log R, the head’s remaining error over the read-off’s | +0.883 (+0.726–+1.047) — Unseen question. R 2.42; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · light sleep · implicit name · three seeds | log R, the head’s remaining error over the read-off’s | +2.267 (+2.072–+2.466) — Unseen question. R 9.65; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · label template against the read-off (TPL / read-off) · light sleep · explicit name · three seeds | log R, the head’s remaining error over the read-off’s | +1.519 (+1.343–+1.698) — Unseen question. R 4.57; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · description against the read-off (DESC / read-off) · asleep · new descriptions · three seeds | log R, the head’s remaining error over the read-off’s | +3.444 (+2.913–+3.956) — Unseen question. R 31.30; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · description against the read-off (DESC / read-off) · N1, N2 or N3 sleep · new descriptions · three seeds | log R, the head’s remaining error over the read-off’s | +1.299 (+0.793–+1.744) — Unseen question. R 3.67; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P3 · description against the read-off (DESC / read-off) · light sleep · new descriptions · three seeds | log R, the head’s remaining error over the read-off’s | +1.417 (+0.903–+1.856) — Unseen question. R 4.12; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P5 · label template − shuffled templates (TPL − SHUF) · unseen questions, implicit names · three seeds | Paired difference in AUROC | −0.378 (−0.501–−0.253) — Unseen question. SHUF higher; difference only: no margin; headroom gate passed. | 100 | | [CBraMod](https://bci.report/methods/cbramod/) | P5 · label template − shuffled templates (TPL − SHUF) · unseen questions, explicit names · three seeds | Paired difference in AUROC | +0.629 (+0.603–+0.654) — Unseen question. TPL higher; difference only: no margin; headroom gate passed. | 100 | **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. **Limits:** 'Unseen' means unseen by the EEG head; the text encoder has read these words, and every token of a held-out frequency also occurs in the training texts. Held-out sleep questions are combinations of seen stages ('asleep' is the complement of one), reported as compositional beside the read-off. Rewording and unseen results are separate rows and never pooled. 35 pre-declared comparisons, not corrected for multiplicity. The pre-run permutation canary failed twice (revisions 3 and 4) and was revised by the owner, who overrode revision 4's "no further revision"; revision 5 passed. Both failed records are kept and disclosed. [Every limitation of route 3 →](https://bci.report/topics/questions-in-language/#methods-and-limits) ### [WBCIC-SHU motor imagery dataset](https://bci.report/datasets/wbcic-shu/) · Later sessions · two-class motor imagery, session 1 to session 3 Source page: [Later sessions](https://bci.report/topics/later-sessions/#wbcic-shu) · Chance level 50.0% · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 67.7% (64.8%–70.5%) | 51 | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Minus the relative spectral ridge, the same people and trials | Paired difference | +13.8 pp (+11.3 pp–+16.4 pp) — A comparison of two fixed pipelines on reused test trials; whether pretraining caused the gain, and whether this checkpoint saw WBCIC-SHU in pretraining, are open. | 51 | ### [WBCIC-SHU motor imagery dataset](https://bci.report/datasets/wbcic-shu/) · Later sessions · three-class motor imagery, session 1 to session 3 Source page: [Later sessions](https://bci.report/topics/later-sessions/#wbcic-shu) · Chance level 33.3% · [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Session 1 → session 3 · no session-3 labels | Balanced accuracy | 51.6% (45.2%–58.3%) | 11 | | [CBraMod · frozen encoder, session-1 ridge readout](https://bci.report/methods/cbramod/) | Minus the relative spectral ridge, the same people and trials | Paired difference | +14.0 pp (+7.5 pp–+20.8 pp) — A comparison of two fixed pipelines on reused test trials; whether pretraining caused the gain, and whether this checkpoint saw WBCIC-SHU in pretraining, are open. | 11 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2412.07236) | ## Cite this page BCI Report (2026). *CBraMod: results on public EEG datasets*. https://bci.report/methods/cbramod/ Figures from releases [`questions-in-language-update-20261008`](https://bci.report/releases/#questions-in-language-update-20261008) (2026-10-08), [`later-sessions-update-20261008`](https://bci.report/releases/#later-sessions-update-20261008) (2026-10-08), [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007) (2026-10-07), [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/cbramod/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # CCA: results on public EEG datasets Classical method CCA-based SSVEP detection, proposed by Lin, Zhang, Wu and Gao (Tsinghua University; IEEE Transactions on Biomedical Engineering, 2006), computes the canonical correlation between multichannel EEG and sine–cosine reference signals at each candidate stimulus frequency and its harmonics. The frequency with the largest correlation is selected; the references are generated from the stimulus frequencies, not learned from recorded trials. Measured on: [BETA](https://bci.report/datasets/beta/); [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/); [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) · Protocols: [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) **Also known as** CCA · Standard CCA · canonical correlation analysis · Lin et al. 2006 **Reference** [Lin Z, Zhang C, Wu W, Gao X. Frequency recognition based on canonical correlation analysis for SSVEP-based BCIs. IEEE Transactions on Biomedical Engineering 53(12):2610–2614, 2006. ↗](https://doi.org/10.1109/TBME.2006.886577) Directory status: Evaluated · [Implementation used here ↗](https://scikit-learn.org/stable/) · Description sources [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/17152442/) · [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/17549911/) · [doi.org](https://doi.org/10.1109/TBME.2006.886577) · [doi.org](https://doi.org/10.1109/tbme.2006.889197) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC5783827/) · [github.com](https://github.com/mnakanishi/TRCA-SSVEP/blob/master/src/test_fbcca.m) · [iopscience.iop.org](https://iopscience.iop.org/article/10.1088/1741-2560/12/4/046008) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 63.1% (57.2%–69.0%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 57.6% (51.8%–63.5%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) | Fixed threshold: calibrated confidence at least 0.8 | Trials accepted | 3,176 / 11,200 — 0 of 70 people had nothing accepted. | 70 | | [Standard CCA](https://bci.report/methods/cca/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.3 pp (−0.3 pp–+1.0 pp) — No difference resolved: the interval includes zero. | 70 | **Limits:** Uniform target prior by design; cue-paced lab spelling with no idle state, so no false-activation figure of any kind. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Dry vs. wet sensor transfer Source page: [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CCA · Wet recording · no source training: 81.6% (77.7%–85.1%) - CCA · Dry recording · no source training: 65.5% (60.6%–70.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CCA](https://bci.report/methods/cca/) | Wet recording · no source training | Balanced accuracy | 81.6% (77.7%–85.1%) | 102 | | [CCA](https://bci.report/methods/cca/) | Dry recording · no source training | Balanced accuracy | 65.5% (60.6%–70.2%) | 102 | ### [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Calibration budget · eTRCA and CCA Source page: [How much calibration?](https://bci.report/topics/calibration-budget/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CCA · No calibration · dry: 65.1% (60.0%–70.0%) - CCA · No calibration · wet: 81.1% (77.2%–84.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CCA](https://bci.report/methods/cca/) | No calibration · dry | Balanced accuracy | 65.1% (60.0%–70.0%) | 102 | | [CCA](https://bci.report/methods/cca/) | No calibration · wet | Balanced accuracy | 81.1% (77.2%–84.8%) | 102 | ### [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) · Movement · SSVEP, 5-second windows Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 6 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Author-style CCA · Standing · scalp · 8 ch: 88.7% (80.0%–95.9%) - Author-style CCA · Slow walk · 0.8 m/s · scalp · 8 ch: 82.9% (74.3%–90.4%) - Author-style CCA · Fast walk · 1.6 m/s · scalp · 8 ch: 80.9% (72.0%–88.8%) - Author-style CCA · Standing · ear · 14 ch: 53.1% (47.4%–59.6%) - Author-style CCA · Slow walk · 0.8 m/s · ear · 14 ch: 43.4% (39.6%–47.8%) - Author-style CCA · Fast walk · 1.6 m/s · ear · 14 ch: 39.1% (36.7%–41.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Author-style CCA](https://bci.report/methods/cca/) | Standing · scalp · 8 ch | Balanced accuracy | 88.7% (80.0%–95.9%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 82.9% (74.3%–90.4%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 80.9% (72.0%–88.8%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Standing · ear · 14 ch | Balanced accuracy | 53.1% (47.4%–59.6%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Slow walk · 0.8 m/s · ear · 14 ch | Balanced accuracy | 43.4% (39.6%–47.8%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Fast walk · 1.6 m/s · ear · 14 ch | Balanced accuracy | 39.1% (36.7%–41.8%) | 23 | ### [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) · Movement · SSVEP, 2-second windows Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 3 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Author-style CCA · Standing · scalp · 8 ch: 82.3% (74.6%–89.4%) - Author-style CCA · Slow walk · 0.8 m/s · scalp · 8 ch: 70.7% (62.3%–78.6%) - Author-style CCA · Fast walk · 1.6 m/s · scalp · 8 ch: 68.6% (60.1%–76.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Author-style CCA](https://bci.report/methods/cca/) | Standing · scalp · 8 ch | Balanced accuracy | 82.3% (74.6%–89.4%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 70.7% (62.3%–78.6%) | 23 | | [Author-style CCA](https://bci.report/methods/cca/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 68.6% (60.1%–76.7%) | 23 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Cite this page BCI Report (2026). *CCA: results on public EEG datasets*. https://bci.report/methods/cca/ Figures from releases [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/cca/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # CodeBrain: results on public EEG datasets Foundation model CodeBrain is a pretrained EEG encoder whose paper lists the Temple University EEG corpus (TUEG) as its only pretraining source. BCI Report ran the released checkpoint frozen, pooled as the published CBraMod rows are, and adapted it on EEGMAT, where its LoRA in effect adapts only the attention value projection. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** YjMajy/CodeBrain **Reference** [CodeBrain, arXiv:2506.09110 ↗](https://arxiv.org/abs/2506.09110) Directory status: Evaluated · Description sources [arxiv.org](https://arxiv.org/abs/2506.09110) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 58.8% (54.5%–62.6%) | 10 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 63.0% (59.5%–66.5%) | 36 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - CodeBrain · Frozen encoder + trained head: 66.5% (62.3%–70.6%) - CodeBrain · LoRA rank 4 + head: 67.1% (62.6%–71.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + trained head | Balanced accuracy | 66.5% (62.3%–70.6%) — 7,602 trainable parameters. | 36 | | [CodeBrain](https://bci.report/methods/codebrain/) | LoRA rank 4 + head | Balanced accuracy | 67.1% (62.6%–71.6%) — 58,802 trainable parameters. | 36 | | [CodeBrain](https://bci.report/methods/codebrain/) | LoRA minus frozen + head, same people and folds | Paired difference | +0.6 pp (−0.7 pp–+1.9 pp) — 18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established. | 36 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 11.9% (10.5%–13.3%) | 70 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 11.2% (9.8%–12.7%) | 70 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 50.8% (49.3%–52.2%) | 21 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 54.7% (53.0%–56.6%) | 30 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + ridge head | Balanced accuracy | 70.7% (68.8%–72.7%) | 20 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + linear head | Commands detected ≤3 s | 10 / 60 — 1 of 4 people always abstained. | 4 | | [CodeBrain](https://bci.report/methods/codebrain/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | Weights terms: CodeBrain under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### CodeBrain - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 13.53M - Revision: `YjMajy/CodeBrain @ bef08d2f` - Checkpoint SHA-256: `d9714b8732c9883a04d022ee66254cd578ae1fa27f5458e6ab7f1aa96e9a7352` - Paper: [paper ↗](https://arxiv.org/abs/2506.09110) - Weights licence: Apache-2.0 - Rights review: Permissive weights licence; aggregate scores only, no weights redistributed. - Row footnote: Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2506.09110) | ## Cite this page BCI Report (2026). *CodeBrain: results on public EEG datasets*. https://bci.report/methods/codebrain/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/codebrain/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # CSP+LDA: results on public EEG datasets Classical method CSP+LDA is a classical two-stage pipeline for motor-imagery EEG: common spatial patterns (CSP) learns spatial filters whose output variances differ most between two classes, and linear discriminant analysis (LDA) classifies the log-variances of the filtered signals. Ramoser, Müller-Gerking and Pfurtscheller (2000) applied CSP with log-variance features and a linear classifier to imagined left- and right-hand movement. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** CSP+LDA · Common Spatial Patterns · CSP · Ramoser et al. 2000 **Reference** [Ramoser H, Müller-Gerking J, Pfurtscheller G. Optimal spatial filtering of single trial EEG during imagined hand movement. IEEE Transactions on Rehabilitation Engineering 8(4):441–446, 2000. ↗](https://doi.org/10.1109/86.895946) Directory status: Evaluated · [Implementation used here ↗](https://mne.tools/stable/generated/mne.decoding.CSP.html) · Description sources [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/11204034/) · [doi.org](https://doi.org/10.1109/86.895946) · [mne.tools](https://mne.tools/stable/generated/mne.decoding.CSP.html) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Rest versus right-hand imagery | Balanced accuracy | 47.1% (43.8%–50.2%) | 10 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Serial subtraction versus resting EEG | Balanced accuracy | 68.3% (62.0%–74.3%) — Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. | 36 | ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Visual target versus nontarget events | Balanced accuracy | 50.0% (49.1%–50.8%) — Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. On this ERP protocol it reads band power, not the waveform. | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Reading · target versus nontarget events | Balanced accuracy | 52.7% (51.4%–54.0%) — Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. On this ERP protocol it reads band power, not the waveform. | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [core-matrix-20261106-sleep-scalp.csv](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA (one-vs-rest CSP)](https://bci.report/methods/csp-lda/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 50.1% (46.9%–53.1%) — One-vs-rest CSP: one binary CSP per stage, 30 log-variance features, one shrinkage LDA; broadband as in the published mi-rest recipe; a new code path, not a tuned CSP. | 20 | ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 23.3% | 4 | | [CSP+LDA](https://bci.report/methods/csp-lda/) | Seated motor imagery · cue-gated replay | Idle false activation | 0.0% | 4 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) [TS+LR and MDM against this row on motor imagery & rest →](https://bci.report/methods/riemannian/#vs-csp) ## Cite this page BCI Report (2026). *CSP+LDA: results on public EEG datasets*. https://bci.report/methods/csp-lda/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/csp-lda/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Deep4Net: results on public EEG datasets Compact model Deep4Net is Braindecode's name for the deep ConvNet of Schirrmeister et al. (Human Brain Mapping, 2017). It stacks four convolution–max-pooling blocks with ELU activations; the first block is split into a convolution over time and a spatial filter across all electrodes, and a dense softmax layer gives the class output. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** Deep4Net · Deep ConvNet · DeepConvNet · Schirrmeister et al. 2017 **Reference** [Schirrmeister RT, Springenberg JT, Fiederer LDJ, Glasstetter M, Eggensperger K, Tangermann M, Hutter F, Burgard W, Ball T. Deep learning with convolutional neural networks for EEG decoding and visualization. Human Brain Mapping 38(11):5391–5420, 2017. ↗](https://doi.org/10.1002/hbm.23730) Directory status: Evaluated · [Official code ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) · Description sources [arxiv.org](https://arxiv.org/abs/1703.05051) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC5655781/) · [doi.org](https://doi.org/10.1002/hbm.23730) · [github.com](https://github.com/braindecode/braindecode/blob/v1.5.1/braindecode/models/deep4.py) · [github.com](https://github.com/robintibor/braindecode) · [github.com](https://github.com/TNTLFreiburg/braindecode) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [core-matrix-20261106-mi-rest.csv](https://bci.report/data/core-matrix-20261106-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | Rest versus right-hand imagery | Balanced accuracy | 68.6% (63.3%–74.0%) — One fixed seed and training budget, as the EEGNet cell | 10 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | Serial subtraction versus resting EEG | Balanced accuracy | 61.3% (57.5%–65.2%) — One fixed seed and training budget, as the EEGNet cell | 36 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [core-matrix-20261106-beta-8ch.csv](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 52.8% (47.3%–58.2%) — One fixed seed and training budget, as the EEGNet cell | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [core-matrix-20261106-beta-4ch.csv](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 38.7% (33.7%–43.8%) — One fixed seed and training budget, as the EEGNet cell | 70 | ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | Visual target versus nontarget events | Balanced accuracy | 50.8% (50.0%–51.6%) — Braindecode's automatically scaled Deep4Net for 1-s input (no temporal pooling, 229-sample classifier, 235,752 parameters), not the original Deep4Net configuration. The interval's lower end, 50.04 %, prints as 50.0; it sits 0.04 pp above the 50 % reference. | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | Reading · target versus nontarget events | Balanced accuracy | 57.4% (55.4%–59.5%) — Braindecode's automatically scaled Deep4Net for 1-s input (no temporal pooling, 229-sample classifier, 242,627 parameters), not the original Deep4Net configuration. | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [core-matrix-20261106-sleep-scalp.csv](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 45.2% (42.7%–47.7%) — One fixed seed and training budget, as the EEGNet cell | 20 | ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Deep4Net](https://bci.report/methods/deep4net/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 3.3% | 4 | | [Deep4Net](https://bci.report/methods/deep4net/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% | 4 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Cite this page BCI Report (2026). *Deep4Net: results on public EEG datasets*. https://bci.report/methods/deep4net/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/deep4net/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # eeg-fm-masking: results on public EEG datasets Foundation model eeg-fm-masking is a set of encoders released with a study of masking in EEG pretraining: one small backbone, one corpus (the open-licence subset of the REVE pretraining corpus) and one recipe, pretrained with a masked-autoencoder (MAE) or a joint-embedding predictive (JEPA) objective under different mask geometries. BCI Report ran four of them frozen: the paper-recommended MAE checkpoint (mask radius 9 cm, L = 2) as a matrix row, and three siblings — JEPA with the same geometry, and MAE and JEPA masking one channel — as a masking ablation. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** PierreGtch/eeg-fm-masking_mae_r9cm_L2 · eeg-fm-masking MAE · eeg-fm-masking JEPA **Reference** [eeg-fm-masking, arXiv:2609.33487 ↗](https://arxiv.org/abs/2609.33487) Directory status: Evaluated (frozen probes, masking ablation) · Description sources [arxiv.org](https://arxiv.org/abs/2609.33487) · [huggingface.co](https://huggingface.co/datasets/brain-bzh/reve-dataset) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 68.1% (64.3%–72.3%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 63.2% (58.2%–67.9%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 64.9% (60.3%–69.4%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 63.3% (59.0%–67.4%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 68.1% (64.3%–72.3%) | 10 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 63.2% (58.2%–67.9%) | 10 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 64.9% (60.3%–69.4%) | 10 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 63.3% (59.0%–67.4%) | 10 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 63.6% (59.1%–68.1%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 61.6% (56.4%–66.5%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 64.3% (60.2%–68.2%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 62.5% (58.5%–66.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 63.6% (59.1%–68.1%) | 36 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 61.6% (56.4%–66.5%) | 36 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 64.3% (60.2%–68.2%) | 36 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 62.5% (58.5%–66.2%) | 36 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 54.8% (49.4%–60.0%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 49.6% (44.5%–54.8%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 53.4% (48.1%–58.6%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 51.0% (45.8%–56.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 54.8% (49.4%–60.0%) | 70 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 49.6% (44.5%–54.8%) | 70 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 53.4% (48.1%–58.6%) | 70 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 51.0% (45.8%–56.1%) | 70 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 47.1% (42.1%–52.1%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 43.4% (38.7%–48.3%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 46.7% (41.9%–51.5%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 45.4% (40.6%–50.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 47.1% (42.1%–52.1%) | 70 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 43.4% (38.7%–48.3%) | 70 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 46.7% (41.9%–51.5%) | 70 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 45.4% (40.6%–50.2%) | 70 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 49.5% (48.1%–51.0%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 51.6% (50.4%–52.7%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 51.1% (49.2%–53.0%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 50.0% (48.7%–51.5%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 49.5% (48.1%–51.0%) | 21 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 51.6% (50.4%–52.7%) | 21 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 51.1% (49.2%–53.0%) | 21 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 50.0% (48.7%–51.5%) | 21 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 54.8% (52.7%–56.9%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 55.8% (53.9%–57.6%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 54.9% (53.4%–56.4%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 55.0% (53.4%–56.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 54.8% (52.7%–56.9%) | 30 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 55.8% (53.9%–57.6%) | 30 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 54.9% (53.4%–56.4%) | 30 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 55.0% (53.4%–56.6%) | 30 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - eeg-fm-masking MAE (r=9 cm, L=2) · Frozen encoder + ridge head: 67.9% (65.4%–70.4%) - eeg-fm-masking JEPA (r=9 cm, L=2) · Frozen encoder + ridge head · masking ablation: 70.3% (68.0%–72.7%) - eeg-fm-masking MAE (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 65.2% (63.0%–67.3%) - eeg-fm-masking JEPA (r=one channel, L=1) · Frozen encoder + ridge head · masking ablation: 68.3% (65.7%–70.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head | Balanced accuracy | 67.9% (65.4%–70.4%) | 20 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 70.3% (68.0%–72.7%) | 20 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 65.2% (63.0%–67.3%) | 20 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + ridge head · masking ablation | Balanced accuracy | 68.3% (65.7%–70.8%) | 20 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head | Commands detected ≤3 s | 3 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head | Idle false activations | 0 / 60 | 4 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Commands detected ≤3 s | 12 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Idle false activations | 0 / 60 | 4 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Commands detected ≤3 s | 5 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Idle false activations | 0 / 60 | 4 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Commands detected ≤3 s | 5 / 60 — 2 of 4 people always abstained. | 4 | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) | Frozen encoder + linear head · masking ablation | Idle false activations | 0 / 60 | 4 | Weights terms: eeg-fm-masking under CC-BY-4.0. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### eeg-fm-masking MAE (r=9 cm, L=2) - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 12.69M - Revision: `PierreGtch/eeg-fm-masking_mae_r9cm_L2 @ c861ade2` - Checkpoint SHA-256: `d889abd87a245b92f782246f6fd51b23120208f307bc65b4de587a76ea5ed4f2` - Paper: [paper ↗](https://arxiv.org/abs/2609.33487) - Weights licence: CC-BY-4.0 (cite arXiv:2609.33487) - Licence note: CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). - Rights review: Attribution licence; the paper is cited beside every row. Aggregate scores only. - Row footnote: Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - EEGMAT adaptation: not run by design: eeg-fm-masking is a frozen-probe masking ablation ### eeg-fm-masking JEPA (r=9 cm, L=2) - Panel: Masking ablation: a sibling of the paper-recommended checkpoint, in its own panel - Encoder parameters: 12.69M - Revision: Not in the release: it names only the paper-recommended checkpoint’s revision. - Checkpoint SHA-256: Not in the release: it names only the paper-recommended checkpoint’s revision. - Paper: [paper ↗](https://arxiv.org/abs/2609.33487) - Weights licence: CC-BY-4.0 (cite arXiv:2609.33487) - Licence note: CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). - Rights review: Attribution licence; the paper is cited beside every row. Aggregate scores only. - Row footnote: A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. - EEGMAT adaptation: not run by design: eeg-fm-masking is a frozen-probe masking ablation ### eeg-fm-masking MAE (r=one channel, L=1) - Panel: Masking ablation: a sibling of the paper-recommended checkpoint, in its own panel - Encoder parameters: 12.69M - Revision: Not in the release: it names only the paper-recommended checkpoint’s revision. - Checkpoint SHA-256: Not in the release: it names only the paper-recommended checkpoint’s revision. - Paper: [paper ↗](https://arxiv.org/abs/2609.33487) - Weights licence: CC-BY-4.0 (cite arXiv:2609.33487) - Licence note: CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). - Rights review: Attribution licence; the paper is cited beside every row. Aggregate scores only. - Row footnote: A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. - EEGMAT adaptation: not run by design: eeg-fm-masking is a frozen-probe masking ablation ### eeg-fm-masking JEPA (r=one channel, L=1) - Panel: Masking ablation: a sibling of the paper-recommended checkpoint, in its own panel - Encoder parameters: 12.69M - Revision: Not in the release: it names only the paper-recommended checkpoint’s revision. - Checkpoint SHA-256: Not in the release: it names only the paper-recommended checkpoint’s revision. - Paper: [paper ↗](https://arxiv.org/abs/2609.33487) - Weights licence: CC-BY-4.0 (cite arXiv:2609.33487) - Licence note: CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). - Rights review: Attribution licence; the paper is cited beside every row. Aggregate scores only. - Row footnote: A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. - EEGMAT adaptation: not run by design: eeg-fm-masking is a frozen-probe masking ablation The paper-recommended MAE checkpoint and its JEPA sibling are also read with channels zeroed at the input, with the head kept or refitted: [Channels zeroed on two published encoders, head kept or refitted →](https://bci.report/topics/pilot-studies/#nsp-components-a) ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. ### eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1) | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.33487) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | ## Cite this page BCI Report (2026). *eeg-fm-masking: results on public EEG datasets*. https://bci.report/methods/eeg-fm-masking/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/eeg-fm-masking/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEGMamba: results on public EEG datasets Foundation model EEGMamba is the Mamba-based EEG encoder from the CBraMod group (Wang et al., Neural Networks, 2025), not the unrelated model of the same name on arXiv (2407.20254). Its official code lists five pretraining sources, including TUEG and the Siena Scalp EEG Database; the paywalled paper was not read. BCI Report ran the released checkpoint frozen, pooled as the published CBraMod rows are, and adapted it on EEGMAT. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** weighting666/EEGMamba · Wang et al. 2025 (Neural Networks) **Reference** [Wang et al., EEGMamba, Neural Networks (2025). doi:10.1016/j.neunet.2025.107816 ↗](https://doi.org/10.1016/j.neunet.2025.107816) Directory status: Evaluated · [Official code ↗](https://github.com/wjq-learning/EEGMamba) · Description sources [doi.org](https://doi.org/10.1016/j.neunet.2025.107816) · [github.com](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 67.3% (63.6%–70.6%) | 10 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 60.3% (57.0%–63.4%) | 36 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - EEGMamba · Frozen encoder + trained head: 60.4% (55.7%–65.2%) - EEGMamba · LoRA rank 4 + head: 60.8% (56.3%–65.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + trained head | Balanced accuracy | 60.4% (55.7%–65.2%) — 7,602 trainable parameters. | 36 | | [EEGMamba](https://bci.report/methods/eegmamba/) | LoRA rank 4 + head | Balanced accuracy | 60.8% (56.3%–65.3%) — 90,930 trainable parameters. | 36 | | [EEGMamba](https://bci.report/methods/eegmamba/) | LoRA minus frozen + head, same people and folds | Paired difference | +0.4 pp (−0.7 pp–+1.5 pp) — 18 people improved, 14 got worse, 4 unchanged. The interval includes zero: no change is established. | 36 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 40.2% (35.7%–44.7%) | 70 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 32.7% (28.9%–36.7%) | 70 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 52.5% (51.4%–53.5%) | 21 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 54.4% (52.5%–56.3%) | 30 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + ridge head | Balanced accuracy | 64.1% (61.6%–66.4%) | 20 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + linear head | Commands detected ≤3 s | 2 / 60 — 3 of 4 people always abstained. | 4 | | [EEGMamba](https://bci.report/methods/eegmamba/) | Frozen encoder + linear head | Idle false activations | 0 / 60 | 4 | Weights terms: EEGMamba under MIT. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### EEGMamba - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 3.28M - Revision: `weighting666/EEGMamba @ 0b060d87` - Checkpoint SHA-256: `b452bb29ecf1d6131ba82a50c6e13823ec1d660d9009d013e691d19b2916f4fe` - Paper: [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - Weights licence: MIT - Rights review: Permissive weights licence; aggregate scores only, no weights redistributed. - Row footnote: Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_raw_for_pretraining.py) · [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/preprocessing/preprocessing_for_pretraining/preprocessing_szbd_for_pretraining.py) · [source ↗](https://pubmed.ncbi.nlm.nih.gov/40714477/) | **Note on the check:** EEGMamba's list comes from the official code (five entries, matching the abstract); the paywalled paper was not read, so confidence is medium. ## Cite this page BCI Report (2026). *EEGMamba: results on public EEG datasets*. https://bci.report/methods/eegmamba/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/eegmamba/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEGNet: results on public EEG datasets Compact model EEGNet is a compact convolutional neural network for EEG-based brain–computer interfaces, proposed by Lawhern et al. (U.S. Army Research Laboratory; Journal of Neural Engineering, 2018) and tested on P300, ERN, MRCP and SMR data. It applies a temporal convolution, a depthwise spatial convolution and a separable convolution, followed by a softmax classifier. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/); [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/); [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/); [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** EEGNet · EEGNet-8,2 · EEGNet-4,2 · EEGNetv4 · Lawhern et al. 2018 **Reference** [Lawhern VJ, Solon AJ, Waytowich NR, Gordon SM, Hung CP, Lance BJ. EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces. Journal of Neural Engineering 15(5):056013, 2018. ↗](https://doi.org/10.1088/1741-2552/aace8c) Directory status: Evaluated · [Official code ↗](https://github.com/vlawhern/arl-eegmodels) · [Implementation used here ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) · Description sources [arxiv.org](https://arxiv.org/abs/1611.08024) · [doi.org](https://doi.org/10.1088/1741-2552/aace8c) · [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/29932424/) · [github.com](https://github.com/braindecode/braindecode/blob/v1.5.1/braindecode/models/eegnet.py) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Rest versus right-hand imagery | Balanced accuracy | 71.0% (65.0%–76.3%) | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Reliable decisions · when to decline a decision (crude: ten people) Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 250 / 1,200 — 0 of 10 people had nothing accepted. | 10 | | [EEGNet](https://bci.report/methods/eegnet/) | Learned reject option minus calibrated confidence, same classifier (a secondary contrast) | Error among the 80% most certain, difference | +0.9 pp (−0.4 pp–+2.4 pp) — No difference resolved: the interval includes zero. | 10 | **Limits:** Ten people, two per test fold. Bootstrap intervals are crude; no person-level dispersion is published. Balanced, so no deployment rate. Its learned-reject contrasts are secondary. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Serial subtraction versus resting EEG | Balanced accuracy | 67.6% (62.8%–72.5%) | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 126 / 2,160 — 19 of 36 people had nothing accepted. | 36 | | [EEGNet](https://bci.report/methods/eegnet/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +2.3 pp (+0.6 pp–+4.1 pp) — Difference resolved: the interval excludes zero. | 36 | **Limits:** Balanced by design. Error rates describe a 1:1 mix and are a method comparison, not a deployment rate. Class equals recording, so recording-level differences are part of what any model or selector can use. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 55.8% (50.1%–61.3%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 44.1% (38.6%–49.6%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Fixed threshold: calibrated confidence at least 0.8 | Trials accepted | 1,863 / 11,200 — 10 of 70 people had nothing accepted. | 70 | | [EEGNet](https://bci.report/methods/eegnet/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.8 pp (+0.02 pp–+1.6 pp) — Difference resolved: the interval excludes zero. | 70 | **Limits:** Uniform target prior by design; cue-paced lab spelling with no idle state, so no false-activation figure of any kind. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Visual target versus nontarget events | Balanced accuracy | 53.3% (51.7%–54.9%) | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Reading · target versus nontarget events | Balanced accuracy | 61.4% (58.9%–64.0%) | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 56.5% (54.9%–58.3%) | 20 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · One model, several questions · sleep, the full EESM19 release (crude: 20 people, one seed) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · sleep stage, five stages (SL-A) · trained from scratch, one seed | Balanced accuracy | 63.2% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · sleep stage, five stages (SL-A) · trained from scratch, one seed | Balanced accuracy | 66.7% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · sleep stage, five stages (SL-A) · trained from scratch, one seed | Balanced accuracy | 64.7% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +3.47 pp (+1.54–+5.64 pp) — C1 higher; 2 pp margin not met; gate passed. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-A)) | Paired difference, balanced accuracy | −1.47 pp (−2.90–−0.10 pp) — A(SL-A) higher; 2 pp margin not met; gate passed. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · next epoch’s stage differs (SL-E) · trained from scratch, one seed | Balanced accuracy | 53.4% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · next epoch’s stage differs (SL-E) · trained from scratch, one seed | Balanced accuracy | 56.8% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · next epoch’s stage differs (SL-E) · trained from scratch, one seed | Balanced accuracy | 52.1% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +3.38 pp (+2.27–+4.36 pp) — C1 higher; 2 pp margin not met (not read: at floor); at floor: reported, not counted. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-E)) | Paired difference, balanced accuracy | +1.33 pp (+0.42–+2.32 pp) — B-lin higher; B-lin non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · first or second half of the night (SL-F) · trained from scratch, one seed | Balanced accuracy | 61.8% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · first or second half of the night (SL-F) · trained from scratch, one seed | Balanced accuracy | 63.8% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · first or second half of the night (SL-F) · trained from scratch, one seed | Balanced accuracy | 62.5% — One seed: a mean with no interval. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +2.09 pp (+0.38–+3.74 pp) — C1 higher; 2 pp margin not met; gate passed. | 20 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-F)) | Paired difference, balanced accuracy | −0.72 pp (−2.10–+0.69 pp) — No difference shown; 2 pp margin not met: inconclusive at this sample size; gate passed. | 20 | **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). EESM19 full (S13) uses every stored epoch that passes the published core rules at its natural stage mix; it is a different protocol from the site's balanced sleep-scalp subset. SL-E and SL-F are largely predictable from the current stage; no stage-only read-out was computed on EESM19. A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 1.7% | 4 | | [EEGNet](https://bci.report/methods/eegnet/) | Seated motor imagery · cue-gated replay | Idle false activation | 0.0% | 4 | ### [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Dry vs. wet sensor transfer Source page: [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - EEGNet · Trained on wet, tested on wet: 75.0% (70.3%–79.5%) - EEGNet · Trained on wet, tested on dry: 45.6% (40.5%–50.8%) - EEGNet · Trained on dry, tested on wet: 64.0% (59.1%–68.8%) - EEGNet · Trained on dry, tested on dry: 56.8% (51.9%–61.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on wet, tested on wet | Balanced accuracy | 75.0% (70.3%–79.5%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on wet, tested on dry | Balanced accuracy | 45.6% (40.5%–50.8%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on dry, tested on wet | Balanced accuracy | 64.0% (59.1%–68.8%) | 102 | | [EEGNet](https://bci.report/methods/eegnet/) | Trained on dry, tested on dry | Balanced accuracy | 56.8% (51.9%–61.6%) | 102 | ### [OpenBMI motor imagery (Lee et al. 2019)](https://bci.report/datasets/openbmi/) · One model, several questions · motor imagery (route 2) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · Chance level 50.0% · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) **Balanced accuracy, 8 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - EEGNet · fixed heads (B-lin) · imagery or rest (MI-A) · trained from scratch, three seeds: 73.2% (70.2%–76.2%) - EEGNet · shared hidden layer (B-sh) · imagery or rest (MI-A) · trained from scratch, three seeds: 74.3% (71.3%–77.3%) - EEGNet · question-conditioned head (C1) · imagery or rest (MI-A) · trained from scratch, three seeds: 74.6% (71.9%–77.5%) - EEGNet · separate model (A) · imagery or rest (MI-A) · trained from scratch, three seeds: 74.2% (71.3%–76.9%) - EEGNet · fixed heads (B-lin) · which hand (MI-B) · trained from scratch, three seeds: 67.5% (64.6%–70.5%) - EEGNet · shared hidden layer (B-sh) · which hand (MI-B) · trained from scratch, three seeds: 67.0% (64.1%–69.9%) - EEGNet · question-conditioned head (C1) · which hand (MI-B) · trained from scratch, three seeds: 67.4% (64.4%–70.5%) - EEGNet · separate model (A) · which hand (MI-B) · trained from scratch, three seeds: 69.0% (65.9%–72.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 73.2% (70.2%–76.2%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 74.3% (71.3%–77.3%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 74.6% (71.9%–77.5%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · imagery or rest (MI-A) · trained from scratch, three seeds | Balanced accuracy | 74.2% (71.3%–76.9%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | imagery or rest (MI-A) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +1.48 pp (+0.09–+2.89 pp) — C1 higher; 2 pp margin not met; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | imagery or rest (MI-A) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | +0.31 pp (−0.91–+1.54 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | imagery or rest (MI-A) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(MI-A)) | Paired difference, balanced accuracy | −1.00 pp (−2.21–+0.19 pp) — No difference shown; 2 pp margin not met: inconclusive at this sample size; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 67.5% (64.6%–70.5%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 67.0% (64.1%–69.9%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 67.4% (64.4%–70.5%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · which hand (MI-B) · trained from scratch, three seeds | Balanced accuracy | 69.0% (65.9%–72.0%) | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | which hand (MI-B) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | −0.09 pp (−1.24–+1.04 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | which hand (MI-B) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | +0.45 pp (−0.60–+1.34 pp) — No difference shown; equivalent within ±2 pp; gate passed. | 51 | | [EEGNet](https://bci.report/methods/eegnet/) | which hand (MI-B) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(MI-B)) | Paired difference, balanced accuracy | −1.43 pp (−2.57–−0.39 pp) — A(MI-B) higher; 2 pp margin not met; gate passed. | 51 | **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. The OpenBMI numbers here use route 2's own 2-s windows, label-free high-pass and person-disjoint folds; they are not comparable with the site's OpenBMI cross-session calibration figures. A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) ### [BOAS · Bitbrain Open Access Sleep dataset](https://bci.report/datasets/boas/) · One model, several questions · sleep (route 2) Source page: [One model, several questions](https://bci.report/topics/shared-encoder/) · [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) **Balanced accuracy, 12 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - EEGNet · fixed heads (B-lin) · sleep stage, five stages (SL-A) · trained from scratch, three seeds: 34.4% (32.1%–37.0%) - EEGNet · shared hidden layer (B-sh) · sleep stage, five stages (SL-A) · trained from scratch, three seeds: 42.9% (40.1%–45.7%) - EEGNet · question-conditioned head (C1) · sleep stage, five stages (SL-A) · trained from scratch, three seeds: 41.4% (38.6%–44.2%) - EEGNet · separate model (A) · sleep stage, five stages (SL-A) · trained from scratch, three seeds: 35.3% (32.9%–37.8%) - EEGNet · fixed heads (B-lin) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds: 54.4% (53.5%–55.4%) - EEGNet · shared hidden layer (B-sh) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds: 58.9% (57.5%–60.2%) - EEGNet · question-conditioned head (C1) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds: 57.5% (56.0%–58.8%) - EEGNet · separate model (A) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds: 53.2% (52.2%–54.3%) - EEGNet · fixed heads (B-lin) · first or second half of the night (SL-F) · trained from scratch, three seeds: 54.2% (53.0%–55.6%) - EEGNet · shared hidden layer (B-sh) · first or second half of the night (SL-F) · trained from scratch, three seeds: 55.8% (54.2%–57.4%) - EEGNet · question-conditioned head (C1) · first or second half of the night (SL-F) · trained from scratch, three seeds: 55.6% (54.2%–57.1%) - EEGNet · separate model (A) · first or second half of the night (SL-F) · trained from scratch, three seeds: 52.2% (51.0%–53.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 34.4% (32.1%–37.0%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 42.9% (40.1%–45.7%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 41.4% (38.6%–44.2%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · sleep stage, five stages (SL-A) · trained from scratch, three seeds | Balanced accuracy | 35.3% (32.9%–37.8%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +6.96 pp (+4.92–+9.09 pp) — C1 higher; 2 pp margin not met; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | −1.47 pp (−3.79–+0.94 pp) — No difference shown; B-sh non-inferior at 2 pp; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | sleep stage, five stages (SL-A) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-A)) | Paired difference, balanced accuracy | −0.84 pp (−2.73–+0.91 pp) — No difference shown; 2 pp margin not met: inconclusive at this sample size; gate passed. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 54.4% (53.5%–55.4%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 58.9% (57.5%–60.2%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 57.5% (56.0%–58.8%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · next epoch’s stage differs (SL-E) · trained from scratch, three seeds | Balanced accuracy | 53.2% (52.2%–54.3%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +3.02 pp (+1.55–+4.35 pp) — C1 higher; 2 pp margin not met (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | −1.41 pp (−2.40–−0.50 pp) — B-sh higher; B-sh non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | next epoch’s stage differs (SL-E) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-E)) | Paired difference, balanced accuracy | +1.20 pp (+0.07–+2.44 pp) — B-lin higher; B-lin non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | fixed heads (B-lin) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 54.2% (53.0%–55.6%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | shared hidden layer (B-sh) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 55.8% (54.2%–57.4%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | question-conditioned head (C1) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 55.6% (54.2%–57.1%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | separate model (A) · first or second half of the night (SL-F) · trained from scratch, three seeds | Balanced accuracy | 52.2% (51.0%–53.7%) | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · conditioned head minus fixed heads (C1 − B-lin) | Paired difference, balanced accuracy | +1.42 pp (−0.25–+2.90 pp) — No difference shown; 2 pp margin not met (not read: at floor): inconclusive at this sample size; at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · conditioned head minus the same layer not told the question (C1 − B-sh) | Paired difference, balanced accuracy | −0.14 pp (−1.41–+1.15 pp) — No difference shown; equivalent within ±2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | first or second half of the night (SL-F) · fixed heads on one shared small-CNN trunk minus separate models (B-lin − A(SL-F)) | Paired difference, balanced accuracy | +2.01 pp (+0.54–+3.89 pp) — B-lin higher; B-lin non-inferior at 2 pp (not read: at floor); at floor: reported, not counted. | 100 | | [EEGNet](https://bci.report/methods/eegnet/) | awake or asleep (SL-B) · its own model, A(SL-B), against reading it off A(SL-A) | log R, the dedicated model’s remaining error over the read-out’s | −0.380 (−0.583–−0.203) — Dedicated model leaves less error; 20% margin not met; gate passed. | 100 | **BOAS is published with three stated gaps (owner decision, 7 October 2026):** - The consent statement does not say whether participants agreed to public sharing or secondary use. - The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. - The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. **Credit:** Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. **Limits:** The questions are label-backed classification targets from one dataset each, asked by identifier. They are not language prompts and say nothing about unseen questions (route 3). Per-person balanced accuracy averages people equally; it compares arms and is not a deployment error rate. SL-E and SL-F are largely predictable from the current stage: read off the true current stage alone, they reach 61.9% and 59.2% balanced accuracy (100 people; descriptive). A contrast shows a difference when its paired 95% interval excludes zero; it is equivalent when the whole interval lies within ±2 pp, the margin fixed before any result, and an arm is non-inferior when the interval rules out its being 2 pp or more behind. No multiplicity correction: across route 2’s 21 primary comparisons and its secondary ones, about one in 20 comparisons with no true difference may show one by chance. [Every limitation of route 2 →](https://bci.report/topics/shared-encoder/#methods-and-limits) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Cite this page BCI Report (2026). *EEGNet: results on public EEG datasets*. https://bci.report/methods/eegnet/ Figures from releases [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007) (2026-10-07), [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/eegnet/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # EEGPT: results on public EEG datasets Foundation model EEGPT is a pretrained EEG Transformer (NeurIPS 2024). Its paper pretrains it on five public datasets — PhysioMI, HGD, TSU, SEED and M3CV — at 58 electrodes, 256 Hz and 4-second windows, and summarises the channels of each 0.25-second patch in four summary tokens. BCI Report ran the released checkpoint frozen on the eight core protocols, in the same runs as the LaBraM rows and, except on idle, the CBraMod rows, with the site’s fixed readout on those tokens in place of the authors’ learnable spatial filter and linear head. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** EEGPT · BINE022/EEGPT · eegpt_mcae_58chs_4s_large4E **Reference** [EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals. NeurIPS 2024. ↗](https://doi.org/10.52202/079017-1239) Directory status: Evaluated · [Official code ↗](https://github.com/BINE022/EEGPT) · Description sources [papers.neurips.cc](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [doi.org](https://doi.org/10.52202/079017-1239) · [github.com](https://github.com/BINE022/EEGPT/tree/a0e0a8fad729e2ecf4eedb3a81548a6e6d48a705) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · EEGPT, frozen encoder · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 58.3% (54.9%–61.8%) | 10 | ds003810 is recorded at 125 Hz; its windows reach the encoder upsampled to 256 Hz, so they carry nothing above 62.5 Hz. ### [ds003810](https://bci.report/datasets/ds003810/) · Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · pretrained | Balanced accuracy | 58.3% | 10 | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · random initialization (mean) | Balanced accuracy | 54.1% | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · pretrained | Balanced accuracy | 57.0% | 10 | | [EEGPT](https://bci.report/methods/eegpt/) | MI / rest · random initialization (mean) | Balanced accuracy | 53.1% | 10 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · EEGPT, frozen encoder · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 65.7% (60.9%–70.3%) | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · pretrained | Balanced accuracy | 65.7% | 36 | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · random initialization (mean) | Balanced accuracy | 59.2% | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · pretrained | Balanced accuracy | 66.0% | 36 | | [EEGPT](https://bci.report/methods/eegpt/) | Mental workload · random initialization (mean) | Balanced accuracy | 58.6% | 36 | ### [BETA](https://bci.report/datasets/beta/) · EEGPT, frozen encoder · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#eegpt) · Chance level 2.5% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 20.3% (18.2%–22.6%) | 70 | PO5 and PO6 use channel embeddings that EEGPT’s pretraining never trained: they are two of the four names in its 62-name channel dictionary (AF7, AF8, PO5, PO6) that its 58-electrode pretraining montage leaves out. The other six channels here, and all four of the 4-channel subset, use trained embeddings. ### [BETA](https://bci.report/datasets/beta/) · EEGPT, frozen encoder · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#eegpt) · Chance level 2.5% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 17.3% (15.5%–19.2%) | 70 | ### [ds006593](https://bci.report/datasets/ds006593/) · EEGPT, frozen encoder · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 54.0% (51.5%–56.6%) | 21 | On the P300 and semantic targets, one-second windows ran as four real 0.25-second patches: the patch grid was set to four after the encoder was built, with no padding. EEGPT’s encoder treats each time patch on its own, so this is numerically the same as building it for 256-sample inputs, as the Orthosis pilot built it for its 128-sample windows. ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · EEGPT, frozen encoder · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#eegpt) · Chance level 50.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 54.6% (53.3%–55.9%) | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · EEGPT, frozen encoder · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#eegpt) · Chance level 20.0% · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + ridge head | Balanced accuracy | 75.5% (73.9%–77.1%) | 20 | ### [ds005342](https://bci.report/datasets/ds005342/) · EEGPT, frozen encoder · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#eegpt) · [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Commands detected ≤3 s | 7 / 60 — 1 of 4 people always abstained. | 4 | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Idle false activations | 2 / 60 | 4 | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Window balanced accuracy | 57.2% | 4 | | [EEGPT](https://bci.report/methods/eegpt/) | Frozen encoder + linear head | Window AUROC | 0.780 | 4 | Same run as the published LaBraM idle row; the published CBraMod idle row comes from another run, so a comparison with it is across runs. The window balanced accuracy and AUROC were computed from this run’s saved probabilities with the definitions of the further foundation-encoder rows; the run itself recorded the counts. [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. ### EEGPT - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 25.29M - Revision: `BINE022/EEGPT @ a0e0a8fad729e2ecf4eedb3a81548a6e6d48a705` - Checkpoint SHA-256: `eegpt_mcae_58chs_4s_large4E.ckpt` `9d63ebc80bba1ad5e3277df4eecfa05d2745e6243a12ff743c4c0a93930dfac0` - Paper: [EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals (NeurIPS 2024) ↗](https://doi.org/10.52202/079017-1239) - Code licence: Apache-2.0 (the repository’s source code only) - Weights licence: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - Weights: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - Rights review: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - Input: the published 250 Hz microvolt windows, resampled to 256 Hz and divided by 1,000 (millivolts); only the recorded channels, each as its channel token (T3, T4, T5 and T6 named T7, T8, P7 and P8); no interpolation, padding, zero-filled channels or average reference - Pooling: the four summary tokens averaged over time patches and concatenated: 2,048 values per window - Row footnote: Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. - Pilot comparison: [Frozen EEGPT against matched classical features on the Orthosis error-detection dataset →](https://bci.report/topics/pilot-studies/#orthosis) ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-12 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://physionet.org/content/eegmat/1.0.0/) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://doi.org/10.3389/fnins.2020.00627) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-12) | [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) · [source ↗](https://github.com/BINE022/EEGPT/tree/ea7282f698c406cd423ccbedbe9960853761144c/datasets/pretrain) · [source ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | **Pretraining corpus:** Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. **Note on the check:** For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated. ## Cite this page BCI Report (2026). *EEGPT: results on public EEG datasets*. https://bci.report/methods/eegpt/ Figures from release [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/eegpt/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ERP-FM: results on public EEG datasets Foundation model ERP-FM is an EEG foundation model pretrained on event-related potential recordings only: its paper lists the task-level ERP datasets it used. On this site the two ERP protocols, P300 target and semantic target, are in its design; the other six are negative controls outside it, and the ERP windows here have no pre-stimulus baseline, so these results test ERP-FM outside its own setting. The weights are CC BY-NC-SA 4.0, for non-commercial use. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** ERP-FM Base **Reference** [ERP-FM, arXiv:2609.32796 ↗](https://arxiv.org/abs/2609.32796) Directory status: Evaluated (frozen probes) · [Official code ↗](https://github.com/DL4mHealth/ERP-FM) · Description sources [arxiv.org](https://arxiv.org/abs/2609.32796) · [github.com](https://github.com/DL4mHealth/ERP-FM) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 54.8% (51.8%–57.8%) | 10 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 61.1% (57.3%–65.1%) | 36 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 18.4% (16.3%–20.5%) | 70 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 20.9% (18.4%–23.5%) | 70 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 51.2% (49.3%–53.2%) | 21 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 53.3% (51.7%–54.8%) | 30 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + ridge head | Balanced accuracy | 71.0% (69.5%–72.6%) | 20 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + linear head | Commands detected ≤3 s | 3 / 60 — 2 of 4 people always abstained. | 4 | | [ERP-FM Base](https://bci.report/methods/erp-fm/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | Weights terms: ERP-FM under CC BY-NC-SA 4.0, non-commercial. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### ERP-FM Base - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 2.02M - Revision: `Google Drive file 1IXYnCOpDmY6raIn0M8_ggBOVey_A27l1 (no revision id; sha256 is the identity)` - Checkpoint SHA-256: `856564c4208e9cfe30cfeb5052038071c4951b7078de2bc9e461dead7eb511d4` - Paper: [paper ↗](https://arxiv.org/abs/2609.32796) - Weights licence: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) - Licence note: CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. - Rights review: Non-commercial share-alike weights. This site is personal noncommercial research and publishes aggregate scores only, with attribution; no weights are redistributed. - Row footnote: In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - EEGMAT adaptation: not run by design: ERP-FM is frozen probes only and EEGMAT arithmetic is outside the ERP design ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2609.32796) · [source ↗](https://github.com/DL4mHealth/ERP-FM#datasets) | ## Cite this page BCI Report (2026). *ERP-FM: results on public EEG datasets*. https://bci.report/methods/erp-fm/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/erp-fm/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # eTRCA: results on public EEG datasets Methods Ensemble task-related component analysis (ensemble TRCA) is an SSVEP detection method proposed by Nakanishi et al. (IEEE Transactions on Biomedical Engineering, 2018). TRCA learns spatial filters that maximize the reproducibility of time-locked responses across calibration trials; the ensemble version combines the filters learned for every stimulus frequency and correlates a filtered test trial with each class's averaged calibration template. Measured on: [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) **Also known as** ensemble TRCA · ensemble task-related component analysis · TRCA · task-related component analysis · TRCA-SSVEP · Nakanishi et al. 2018 **Reference** [Nakanishi M, Wang Y, Chen X, Wang Y-T, Gao X, Jung T-P. Enhancing detection of SSVEPs for a high-speed brain speller using task-related component analysis. IEEE Transactions on Biomedical Engineering 65(1):104–112, 2018. ↗](https://doi.org/10.1109/TBME.2017.2694818) [Official code ↗](https://github.com/mnakanishi/TRCA-SSVEP) · Description sources [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/28436836/) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC5783827/) · [doi.org](https://doi.org/10.1109/TBME.2017.2694818) · [github.com](https://github.com/mnakanishi/TRCA-SSVEP) · [github.com](https://github.com/mnakanishi/TRCA-SSVEP/blob/master/README.md) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Calibration budget · eTRCA and CCA Source page: [How much calibration?](https://bci.report/topics/calibration-budget/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - Single-band eTRCA · 24 calibration trials · dry: 12.4% (10.4%–15.0%) - Single-band eTRCA · 24 calibration trials · wet: 39.8% (34.1%–45.8%) - Single-band eTRCA · 48 calibration trials · dry: 15.1% (12.6%–18.2%) - Single-band eTRCA · 48 calibration trials · wet: 64.0% (57.1%–70.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 24 calibration trials · dry | Balanced accuracy | 12.4% (10.4%–15.0%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 24 calibration trials · wet | Balanced accuracy | 39.8% (34.1%–45.8%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 48 calibration trials · dry | Balanced accuracy | 15.1% (12.6%–18.2%) | 102 | | [Single-band eTRCA](https://bci.report/methods/etrca/) | 48 calibration trials · wet | Balanced accuracy | 64.0% (57.1%–70.7%) | 102 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Cite this page BCI Report (2026). *eTRCA: results on public EEG datasets*. https://bci.report/methods/etrca/ Figures from release [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/etrca/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # FBCCA: results on public EEG datasets Classical method Filter bank canonical correlation analysis (FBCCA), proposed by Chen, Wang, Gao, Jung and Gao (Journal of Neural Engineering, 2015), extends CCA-based SSVEP detection: a filter bank splits the EEG into several sub-bands, CCA against sine–cosine references is computed in each, and a weighted combination of the sub-band correlations selects the target frequency. It was introduced with a 40-target SSVEP speller. Measured on: [BETA](https://bci.report/datasets/beta/); [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Protocols: [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) **Also known as** FBCCA · filter bank CCA · filter bank canonical correlation analysis · Chen et al. 2015 **Reference** [Chen X, Wang Y, Gao S, Jung T-P, Gao X. Filter bank canonical correlation analysis for implementing a high-speed SSVEP-based brain–computer interface. Journal of Neural Engineering 12(4):046008, 2015. ↗](https://doi.org/10.1088/1741-2560/12/4/046008) Description sources [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/26035476/) · [doi.org](https://doi.org/10.1088/1741-2560/12/4/046008) · [iopscience.iop.org](https://iopscience.iop.org/article/10.1088/1741-2560/12/4/046008) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC5783827/) · [github.com](https://github.com/mnakanishi/TRCA-SSVEP/blob/master/src/test_fbcca.m) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [core-matrix-20261106-beta-8ch.csv](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Filter-bank CCA (FBCCA-M3-N7-H5)](https://bci.report/methods/fbcca/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 79.1% (75.1%–82.8%) — FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free · 250 Hz | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [core-matrix-20261106-beta-4ch.csv](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [Filter-bank CCA (FBCCA-M3-N7-H5)](https://bci.report/methods/fbcca/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 72.3% (67.5%–76.9%) — FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free · 250 Hz | 70 | ### [Wearable SSVEP BCI dataset (dry and wet electrodes)](https://bci.report/datasets/wearable-ssvep-102/) · Dry vs. wet sensor transfer Source page: [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - FBCCA · Wet recording · no source training: 82.3% (78.7%–85.7%) - FBCCA · Dry recording · no source training: 64.6% (59.5%–69.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [FBCCA](https://bci.report/methods/fbcca/) | Wet recording · no source training | Balanced accuracy | 82.3% (78.7%–85.7%) — Five Butterworth order-4 sub-bands (6–80 to 38–80 Hz), four harmonics | 102 | | [FBCCA](https://bci.report/methods/fbcca/) | Dry recording · no source training | Balanced accuracy | 64.6% (59.5%–69.6%) — Five Butterworth order-4 sub-bands (6–80 to 38–80 Hz), four harmonics | 102 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Cite this page BCI Report (2026). *FBCCA: results on public EEG datasets*. https://bci.report/methods/fbcca/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/fbcca/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # LaBraM: results on public EEG datasets Foundation model LaBraM (Large Brain Model) is an EEG foundation model from Jiang, Zhao and Lu (Shanghai Jiao Tong University; ICLR 2024), described in Base, Large and Huge sizes. A vector-quantized tokenizer turns per-channel EEG patches into discrete codes, and a Transformer is pretrained to predict the codes of masked patches, using about 2,500 hours of EEG from around 20 datasets. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/); [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** LaBraM · Large Brain Model · LaBraM-Base · Labram · Jiang et al. 2024 **Reference** [Jiang W-B, Zhao L-M, Lu B-L. Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI. International Conference on Learning Representations (ICLR), 2024. ↗](https://arxiv.org/abs/2405.18765) Directory status: Evaluated · [Official code ↗](https://github.com/935963004/LaBraM) · Description sources [arxiv.org](https://arxiv.org/abs/2405.18765) · [github.com](https://github.com/935963004/LaBraM) · [github.com](https://github.com/braindecode/braindecode/blob/v1.5.1/braindecode/models/labram.py) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Rest versus right-hand imagery | Balanced accuracy | 53.5% (51.4%–55.4%) | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · pretrained | Balanced accuracy | 53.5% | 10 | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · random initialization (mean) | Balanced accuracy | 55.9% | 10 | ### [ds003810](https://bci.report/datasets/ds003810/) · Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · pretrained | Balanced accuracy | 52.9% | 10 | | [LaBraM](https://bci.report/methods/labram/) | MI / rest · random initialization (mean) | Balanced accuracy | 56.5% | 10 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Serial subtraction versus resting EEG | Balanced accuracy | 64.6% (60.5%–68.6%) | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Pretraining · fixed readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · pretrained | Balanced accuracy | 64.6% | 36 | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · random initialization (mean) | Balanced accuracy | 56.5% | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Pretraining · train-selected readout Source page: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · pretrained | Balanced accuracy | 64.4% | 36 | | [LaBraM](https://bci.report/methods/labram/) | Mental workload · random initialization (mean) | Balanced accuracy | 56.5% | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation · LaBraM on new people, same task, zero labels from the test person Source page: [Which part to update?](https://bci.report/topics/model-adaptation/) · Chance level 50.0% · [adaptation-update.json](https://bci.report/data/adaptation-update.json) **Balanced accuracy, 3 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LaBraM · Head only · encoder frozen: 56.6% (54.2%–59.0%) - LaBraM · Last block + head: 65.7% (61.6%–69.7%) - LaBraM · LoRA rank 4 + head: 64.1% (60.2%–67.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Head only · encoder frozen | Balanced accuracy | 56.6% (54.2%–59.0%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | Last block + head | Balanced accuracy | 65.7% (61.6%–69.7%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | LoRA rank 4 + head | Balanced accuracy | 64.1% (60.2%–67.8%) | 36 | | [LaBraM](https://bci.report/methods/labram/) | LoRA rank 4 + head minus head only | Paired difference | +7.5 pp (+4.8 pp–+10.1 pp) | 36 | | [LaBraM](https://bci.report/methods/labram/) | Last block + head minus head only | Paired difference | +9.1 pp (+6.0 pp–+12.2 pp) | 36 | | [LaBraM](https://bci.report/methods/labram/) | LoRA rank 4 + head minus last block + head | Paired difference | −1.6 pp (−4.4 pp–+1.0 pp) | 36 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Reliable decisions · when to decline a decision Source page: [When not to act](https://bci.report/topics/when-not-to-act/) · [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM · head only, encoder frozen](https://bci.report/methods/labram/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 1 / 2,160 — 35 of 36 people had nothing accepted. | 36 | | [LaBraM · head only, encoder frozen](https://bci.report/methods/labram/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.6 pp (−0.5 pp–+2.7 pp) — No difference resolved: the interval includes zero. | 36 | | [LaBraM · LoRA rank 4 + head](https://bci.report/methods/labram/) | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 44 / 2,160 — 24 of 36 people had nothing accepted. | 36 | | [LaBraM · LoRA rank 4 + head](https://bci.report/methods/labram/) | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +1.9 pp (+0.6 pp–+3.5 pp) — Difference resolved: the interval excludes zero. | 36 | **Limits:** Balanced by design. Error rates describe a 1:1 mix and are a method comparison, not a deployment rate. Class equals recording, so recording-level differences are part of what any model or selector can use. Calibration data are participant-disjoint inner out-of-fold scores (cross-fit). Inner models see about 3/4 of the training people, so they may be less confident than the full model whose scores are tested. This approximation is part of every calibrated figure. Rejecting likely errors on known classes is not detecting unfamiliar input. An uncertainty threshold may not flag an unseen class ([https://arxiv.org/abs/2603.13324](https://arxiv.org/abs/2603.13324)). Route 1 makes no out-of-distribution claim. A contrast is called a difference only when its paired 95% participant-bootstrap interval excludes 0 (no multiplicity correction); otherwise "no difference resolved". No ranking where intervals overlap. Descriptive entries are never called differences. [Every limitation of route 1 →](https://bci.report/topics/when-not-to-act/#rd-limits) ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 10.8% (9.5%–12.2%) | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 12.9% (11.2%–14.6%) | 70 | ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Visual target versus nontarget events | Balanced accuracy | 49.4% (47.7%–51.2%) | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Reading · target versus nontarget events | Balanced accuracy | 53.2% (51.3%–54.9%) | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 71.1% (69.0%–73.2%) | 20 | ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 18.3% | 4 | | [LaBraM](https://bci.report/methods/labram/) | Seated motor imagery · cue-gated replay | Idle false activation | 3.3% | 4 | ### [Mobile BCI dataset (SSVEP and ERP paradigms)](https://bci.report/datasets/mobile-bci/) · Movement · SSVEP, 2-second windows Source page: [On the move](https://bci.report/topics/on-the-move/) · [deployment-topics.json](https://bci.report/data/deployment-topics.json) **Balanced accuracy, 3 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LaBraM · Standing · scalp · 8 ch: 59.5% (53.6%–65.5%) - LaBraM · Slow walk · 0.8 m/s · scalp · 8 ch: 54.7% (49.5%–60.1%) - LaBraM · Fast walk · 1.6 m/s · scalp · 8 ch: 47.8% (43.3%–52.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) | Standing · scalp · 8 ch | Balanced accuracy | 59.5% (53.6%–65.5%) | 23 | | [LaBraM](https://bci.report/methods/labram/) | Slow walk · 0.8 m/s · scalp · 8 ch | Balanced accuracy | 54.7% (49.5%–60.1%) | 23 | | [LaBraM](https://bci.report/methods/labram/) | Fast walk · 1.6 m/s · scalp · 8 ch | Balanced accuracy | 47.8% (43.3%–52.3%) | 23 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2405.18765) | **Note on the check:** SingLEM and LaBraM lists are matched by dataset name and reference; most rows carry no accession number. ## LaBraM at Scale LaBraM at Scale (Pan, Yin, Lu, Li, Zheng and Wang, arXiv 2611.03165) releases encoders with LaBraM’s architecture, pretrained on three collections, Quark, Atom and Molecule, at several sizes. Six were run frozen with the published ridge head: Quark-B, Quark-L, Quark-H and Quark-G, and the Large encoders of Atom and Molecule. Quark’s nine pretraining sources are Alljoined, Game, HMC, SEED, SEED-IV, SEED-V, SEED-VII, THINGS-EEG2 and TUH; Atom’s and Molecule’s lists include five of the core datasets, so those cells carry a flag. ### LaBraM at Scale Quark-B - Parameters: 5.82M - Pooled width: 200 - Checkpoint: `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` · `quark/quark-base.safetensors` - SHA-256: `3ffd4ca8fc7acf63dd89a8fc4be8c5d79666818e62bc61b10693767f0925648e` - Input: divided by 100 - Pooling: the mean of the patch tokens after the pretrained final LayerNorm - Backend: Local Ubuntu / CPU ### LaBraM at Scale Quark-L - Parameters: 46.26M - Pooled width: 400 - Checkpoint: `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` · `quark/quark-large.safetensors` - SHA-256: `8729eea4bb16567b11ce1d86d9e23b485324f4b5e8ad1122964003f7c89cd339` - Input: divided by 100 - Pooling: the mean of the patch tokens after the pretrained final LayerNorm - Backend: Local Ubuntu / CPU ### LaBraM at Scale Quark-H - Parameters: 369.24M - Pooled width: 800 - Checkpoint: `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` · `quark/quark-huge.safetensors` - SHA-256: `246b875445b156d54e6ebb0b97f90dbd6890e848e8255290e7d9c7c48a2728c5` - Input: divided by 100 - Pooling: the mean of the patch tokens after the pretrained final LayerNorm - Backend: Local Ubuntu / CPU ### LaBraM at Scale Quark-G - Parameters: 1967.59M - Pooled width: 1,600 - Checkpoint: `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` · `quark/quark-giant.safetensors` - SHA-256: `fe75959108ba07f0b435898dc43ceefca1de6b92b4f77c0486a6bec670ed835e` - Backend: Local Ubuntu / CUDA ### LaBraM at Scale Atom-L - Parameters: 46.26M - Pooled width: 400 - Checkpoint: `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` · `atom/atom-large.safetensors` - SHA-256: `a28e97f7bd73bd7ff55f3b5d680f48807c335cf3134d1281228a061056aaf52b` - Input: divided by 100 - Pooling: the mean of the patch tokens after the pretrained final LayerNorm - Backend: Local Ubuntu / CPU ### LaBraM at Scale Molecule-L - Parameters: 46.26M - Pooled width: 400 - Checkpoint: `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` · `molecule/molecule-large.safetensors` - SHA-256: `aa590160283a5ba5a081ee89f1535f1ade584ffc27455d409ab277be3b4ac8d0` - Input: a per-channel z-score over each segment, clipped at ±15 - Pooling: the mean of the patch tokens after the pretrained final LayerNorm - Backend: Local Ubuntu / CPU Their cells, the Quark ladder and [the Quark-G step](https://bci.report/topics/does-pretraining-help/#fc-quark-g-step) are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#more-checkpoints)”; each cell is also on its protocol page. Terms: code and weights under MIT, ungated (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27; code AlbertMentat/LaBraM-at-Scale @ faa387d9). Cite [arXiv 2611.03165](https://arxiv.org/abs/2611.03165) and, for the architecture, LaBraM ([arXiv 2405.18765](https://arxiv.org/abs/2405.18765)). No weights are distributed. [foundation-checkpoints-update.json](https://bci.report/data/foundation-checkpoints-update.json) · [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Cite this page BCI Report (2026). *LaBraM: results on public EEG datasets*. https://bci.report/methods/labram/ Figures from releases [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`reliable-decisions-update-20261004`](https://bci.report/releases/#reliable-decisions-update-20261004) (2026-10-04), [`adaptation-update-20261001`](https://bci.report/releases/#adaptation-update-20261001) (2026-10-01), [`deployment-topics-20260920-v1`](https://bci.report/releases/#deployment-topics-20260920-v1) (2026-09-20), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/labram/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # LUNA: results on public EEG datasets Foundation model LUNA is a pretrained EEG encoder released in Base, Large and Huge sizes; its paper lists the Temple University EEG corpus (TUEG) and the Siena Scalp EEG Database as its pretraining data. BCI Report ran Base and Large frozen. The input is resampled to 256 Hz and cut to whole patches, and the one- and two-second windows and four- to eight-channel layouts here are outside the regime its authors tested. The weights are CC BY-ND 4.0. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** LUNA Base · LUNA Large · PulpBio/LUNA **Reference** [LUNA, arXiv:2510.22257 ↗](https://arxiv.org/abs/2510.22257) Directory status: Evaluated · [Official code ↗](https://huggingface.co/PulpBio/LUNA) · Description sources [arxiv.org](https://arxiv.org/abs/2510.22257) · [huggingface.co](https://huggingface.co/PulpBio/LUNA) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 52.6% (50.2%–54.9%) - LUNA Large: 49.9% (46.6%–53.5%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 52.6% (50.2%–54.9%) | 10 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 49.9% (46.6%–53.5%) | 10 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 60.7% (56.9%–64.4%) - LUNA Large: 57.4% (54.3%–60.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 60.7% (56.9%–64.4%) | 36 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 57.4% (54.3%–60.6%) | 36 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base · Frozen encoder + trained head: 58.2% (54.5%–61.7%) - LUNA Base · LoRA rank 4 + head: 60.0% (55.9%–63.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + trained head | Balanced accuracy | 58.2% (54.5%–61.7%) — 514 trainable parameters. | 36 | | [LUNA Base](https://bci.report/methods/luna/) | LoRA rank 4 + head | Balanced accuracy | 60.0% (55.9%–63.9%) — 37,378 trainable parameters. | 36 | | [LUNA Base](https://bci.report/methods/luna/) | LoRA minus frozen + head, same people and folds | Paired difference | +1.8 pp (+0.1 pp–+3.3 pp) — 23 people improved, 11 got worse, 2 unchanged. The interval excludes zero. | 36 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 8.6% (7.7%–9.5%) - LUNA Large: 10.0% (8.9%–11.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 8.6% (7.7%–9.5%) | 70 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 10.0% (8.9%–11.1%) | 70 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 9.0% (8.1%–10.1%) - LUNA Large: 11.2% (9.8%–12.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 9.0% (8.1%–10.1%) | 70 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 11.2% (9.8%–12.6%) | 70 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 48.7% (46.3%–51.1%) - LUNA Large: 48.8% (46.9%–50.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 48.7% (46.3%–51.1%) | 21 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 48.8% (46.9%–50.8%) | 21 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 56.2% (54.2%–58.2%) - LUNA Large: 57.1% (55.3%–58.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 56.2% (54.2%–58.2%) | 30 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 57.1% (55.3%–58.9%) | 30 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - LUNA Base: 67.5% (65.3%–69.5%) - LUNA Large: 72.4% (70.4%–74.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 67.5% (65.3%–69.5%) | 20 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + ridge head | Balanced accuracy | 72.4% (70.4%–74.0%) | 20 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + linear head | Commands detected ≤3 s | 4 / 60 — 2 of 4 people always abstained. | 4 | | [LUNA Base](https://bci.report/methods/luna/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + linear head | Commands detected ≤3 s | 5 / 60 — 2 of 4 people always abstained. | 4 | | [LUNA Large](https://bci.report/methods/luna/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | Weights terms: LUNA under CC BY-ND 4.0, with no modified weights shared. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### LUNA Base - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 6.57M - Revision: `PulpBio/LUNA @ 999c1af0` - Checkpoint SHA-256: `482839ad9152b6948ff166d2a1638637c54b4d91f1cefadbce6312831ba9b11d` - Paper: [paper ↗](https://arxiv.org/abs/2510.22257) - Weights licence: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) - Licence note: CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. - Rights review: Owner-approved on 2026-10-04 for internal probing and adaptation with aggregate scores only. No modified weights, LoRA or adapter deltas are shared; no endorsement by the LUNA authors is implied, and the LUNA name is not used to suggest an official release. - Row footnote: Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - Notes: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ### LUNA Large - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 40.45M - Revision: `PulpBio/LUNA @ 999c1af0` - Checkpoint SHA-256: `03e0321a1e1a30dc074f7bfe512a901138200494f7694f31ab633e42b105c0e4` - Paper: [paper ↗](https://arxiv.org/abs/2510.22257) - Weights licence: CC BY-ND 4.0 (owner-approved terms as LUNA Base) - Licence note: CC BY-ND 4.0 weights (as LUNA Base). - Rights review: Owner-approved on 2026-10-04 for internal probing and adaptation with aggregate scores only. No modified weights, LoRA or adapter deltas are shared; no endorsement by the LUNA authors is implied, and the LUNA name is not used to suggest an official release. - Row footnote: As LUNA Base. Frozen probes only. - Notes: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - EEGMAT adaptation: not run by design: LUNA Large is frozen probes only in this evaluation’s specification ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. ### LUNA Base · LUNA Large | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.22257) | ## Cite this page BCI Report (2026). *LUNA: results on public EEG datasets*. https://bci.report/methods/luna/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/luna/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Every EEG model and paper we track, and how far we checked it Open register Every EEG model and paper we track, with the furthest step we have checked, what that step showed and the terms of its code, weights and data. ## What the model ladder holds Every number here is a value of [the register file](https://bci.report/data/open-register-20261106.json). - **Rows.** 62 rows on the model ladder: each model and method with a page on this site, the other baselines measured here, the models in the directory, and the other models and papers whose official records we read. By the furthest step checked: 15 at step 1, 15 at step 2, 1 at step 3, 19 at step 4 and 12 at step 5. The dataset ladder has another 488 rows, on [the dataset register](https://bci.report/datasets/register/). The rows here were checked on 16 October 2026, 20 October 2026, 25 October 2026, 28 October 2026 and 6 November 2026. - **Other models and papers.** We read the official records of 42 further models and papers; 2 of them already have a method page here and are merged into that method’s row. ## The model ladder’s steps A model or paper climbs five steps with names of their own, and every badge in the table names the model ladder. A row’s step is the furthest one whose key item is on file (at step 2, the weights licence); an item still missing at its step or an earlier one is listed beside the badge as a gap. The evidence for each step, item by item with its source, is in the JSON file. - Step 1 · Identified: Official record matched: paper, repository or model page - Step 2 · Terms recorded: Licence for each part (code, weights, data); file listing read - Step 3 · Files checked: Downloaded, hash-checked, restored or smoke-tested - Step 4 · Scored here: A result computed here and published on a page - Step 5 · In the matrix: Comparable on the core protocol pages ### Rules - Rows state facts. - Every row has a correction route. Gated data and suspected re-uploads have rules of their own, on [the dataset register](https://bci.report/datasets/register/#levels). To correct a row, use the mail link at its end: the subject names the row’s id. How corrections are handled is on [the data-use page](https://bci.report/data-use/#register-corrections). ## Every row Each row links its page on this site or the model’s own records. Terms are given for the code, the weights and the data separately, each with who states them, and a part whose terms we could not find says “not found”. A paper’s venue appears only where the venue’s proceedings or the paper’s camera-ready version show it accepted, with the date we checked; where the proceedings are online, the venue links them. Whether a model’s pretraining data include the data it is scored on appears as its pretraining exposure. | Model or paper | Kind · signal | Terms by part (code / weights / data) | Checked to | What we found | | --- | --- | --- | --- | --- | | **Models with results on this site** — Models and methods with results on this site; each links its page. | | | | | | [CBraMod](https://bci.report/methods/cbramod/) — [Paper ↗](https://arxiv.org/abs/2412.07236) · [Code ↗](https://github.com/wjq-learning/CBraMod) — Also: Criss-Cross Brain Foundation Model, Wang et al. 2025 · Official weights: weighting666/CBraMod · Weights copies: smoking11/CBraMod, braindecode/cbramod-pretrained, hanhan05/cbramod-pretrained, eeg-telecom-paris/cbramod-base-official, oraculumai/CBraMod-CoreML-Apple | Foundation model · human scalp EEG | Code: `MIT` (this site’s rights record) Weights: `Apache-2.0` (the model card’s metadata) Data: per source (the paper) | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-cbramod) | | [CCA](https://bci.report/methods/cca/) — [Paper ↗](https://doi.org/10.1109/TBME.2006.886577) · [Code ↗](https://scikit-learn.org/stable/) — Also: Standard CCA, canonical correlation analysis, Lin et al. 2006 | Classical method · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-cca) | | [CSP+LDA](https://bci.report/methods/csp-lda/) — [Paper ↗](https://doi.org/10.1109/86.895946) · [Code ↗](https://mne.tools/stable/generated/mne.decoding.CSP.html) — Also: Common Spatial Patterns, CSP, Ramoser et al. 2000 | Classical method · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-csp-lda) | | [Deep4Net](https://bci.report/methods/deep4net/) — [Paper ↗](https://doi.org/10.1002/hbm.23730) · [Code ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) — Also: Deep ConvNet, DeepConvNet, Schirrmeister et al. 2017 | Trained from scratch · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-deep4net) | | [EEGNet](https://bci.report/methods/eegnet/) — [Paper ↗](https://doi.org/10.1088/1741-2552/aace8c) · [Code ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) — Also: EEGNet-8,2, EEGNet-4,2, EEGNetv4, Lawhern et al. 2018 | Trained from scratch · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eegnet) | | [FBCCA](https://bci.report/methods/fbcca/) — [Paper ↗](https://doi.org/10.1088/1741-2560/12/4/046008) — Also: filter bank CCA, filter bank canonical correlation analysis, Chen et al. 2015 | Classical method · human scalp EEG | Code: not found Weights: not applicable Data: per source | Model ladder · step 5 of 5 · In the matrix — gap at step 2: code licence | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-fbcca) | | [LaBraM](https://bci.report/methods/labram/) — [Paper ↗](https://arxiv.org/abs/2405.18765) · [Code ↗](https://github.com/935963004/LaBraM) — Also: Large Brain Model, LaBraM-Base, Labram, Jiang et al. 2024 · Weights copies: braindecode/labram-pretrained, braindecode/Labram-Braindecode, eeg-telecom-paris/labram-base-official | Foundation model · human scalp EEG | Code: `MIT` (this site’s rights record) Weights: `MIT` (this site’s rights record) Data: per source (the paper) | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-labram) | | [TS+LR and MDM](https://bci.report/methods/riemannian/) — [Paper ↗](https://doi.org/10.1109/TBME.2011.2172210) · [Code ↗](https://github.com/pyRiemann/pyRiemann) — Also: TS+LR, tangent space + logistic regression, tangent-space logistic regression, MDM, minimum distance to mean, Barachant et al. 2012 | Classical method · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-riemannian) | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — [Paper ↗](https://doi.org/10.1002/hbm.23730) · [Code ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) — Also: Shallow ConvNet, ShallowConvNet, Schirrmeister et al. 2017 | Trained from scratch · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-shallowfbcspnet) | | [xDAWN+LDA and XDAWNCov+TS+LR](https://bci.report/methods/xdawn/) — [Paper ↗](https://doi.org/10.1109/TBME.2009.2012869) · [Code ↗](https://github.com/pyRiemann/pyRiemann) — Also: xDAWN, XDAWNCov, xDAWN+LDA, XDAWNCov+TS+LR, Rivet et al. 2009 | Classical method · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-xdawn) | | [BrainOmni](https://bci.report/methods/brainomni/) — [Paper ↗](https://arxiv.org/abs/2505.18185) · [Code ↗](https://github.com/OpenTSLab/BrainOmni) — Also: BrainOmni Base, BrainTokenizer · Official weights: OpenTSLab/BrainOmni · Weights copies: braindecode/brainomni-base-pretrained | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-brainomni) | | [CodeBrain](https://bci.report/methods/codebrain/) — [Paper ↗](https://arxiv.org/abs/2506.09110) · [Code ↗](https://github.com/jingyingma01/CodeBrain) — Official weights: YjMajy/CodeBrain | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `Apache-2.0` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-codebrain) | | [eeg-fm-masking](https://bci.report/methods/eeg-fm-masking/) — [Paper ↗](https://arxiv.org/abs/2609.33487) · [Code ↗](https://github.com/PierreGtch/eeg-fm-masking) — Also: eeg-fm-masking MAE, eeg-fm-masking JEPA · Official weights: PierreGtch/eeg-fm-masking_mae_r9cm_L2 | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `CC-BY-4.0` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eeg-fm-masking) | | [EEGMamba](https://bci.report/methods/eegmamba/) — [Paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) · [Code ↗](https://github.com/wjq-learning/EEGMamba) — Also: Wang et al. 2025 (Neural Networks) · Official weights: weighting666/EEGMamba | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eegmamba) | | [EEGPT](https://bci.report/methods/eegpt/) — [Paper ↗](https://doi.org/10.52202/079017-1239) · [Code ↗](https://github.com/BINE022/EEGPT) — Also: BINE022/EEGPT, eegpt_mcae_58chs_4s_large4E · Weights copies: braindecode/eegpt-pretrained, eeg-telecom-paris/eegpt-large-official | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: not found Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here — gap at step 2: weights licence | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eegpt) | | [ERP-FM](https://bci.report/methods/erp-fm/) — [Paper ↗](https://arxiv.org/abs/2609.32796) · [Code ↗](https://github.com/DL4mHealth/ERP-FM) — Also: ERP-FM Base | Foundation model · human scalp EEG | Code: `CC BY-NC-SA 4.0` (the repository’s licence file) Weights: `CC BY-NC-SA 4.0` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-erp-fm) | | [eTRCA](https://bci.report/methods/etrca/) — [Paper ↗](https://doi.org/10.1109/TBME.2017.2694818) · [Code ↗](https://github.com/mnakanishi/TRCA-SSVEP) — Also: ensemble TRCA, ensemble task-related component analysis, TRCA, task-related component analysis, TRCA-SSVEP, Nakanishi et al. 2018 | Classical method · human scalp EEG | Code: not found Weights: not applicable Data: per source | Model ladder · step 4 of 5 · Scored here — gap at step 2: code licence | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-etrca) | | [LUNA](https://bci.report/methods/luna/) — [Paper ↗](https://arxiv.org/abs/2510.22257) · [Code ↗](https://github.com/pulp-bio/BioFoundation) · [Weights ↗](https://huggingface.co/models?search=LUNA) — Also: LUNA Base, LUNA Large · Official weights: PulpBio/LUNA | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `CC BY-ND 4.0` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-luna) | | [REVE](https://bci.report/methods/reve/) — [Paper ↗](https://arxiv.org/abs/2510.21585) · [Code ↗](https://github.com/elouayas/reve_eeg) · [Weights ↗](https://huggingface.co/models?search=reve-base) — Also: REVE Base, REVE Large, REVE Small · Official weights: brain-bzh/reve-base, brain-bzh/reve-large, brain-bzh/reve-small | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `REVE Responsible Use License v1.0` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-reve) | | [SingLEM](https://bci.report/methods/singlem/) — [Paper ↗](https://arxiv.org/abs/2509.17920v3) · [Code ↗](https://github.com/ttlabtuat/SingLEM) — Official weights: ttlabtuat/SingLEM | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-singlem) | | [ST-EEGFormer](https://bci.report/methods/st-eegformer/) — [Code ↗](https://github.com/LiuyinYang1101/STEEGFormer) — Also: ST-EEGFormer Base, ST-EEGFormer Large, STEEGFormer · Weights copies: braindecode/STEEGFormer-base, braindecode/STEEGFormer-large | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-st-eegformer) | | [ZUNA](https://bci.report/methods/zuna/) — [Paper ↗](https://arxiv.org/abs/2607.27308) · [Code ↗](https://github.com/Zyphra/zuna) · [Weights ↗](https://huggingface.co/models?search=ZUNA1.1) — Also: ZUNA 1.1, ZUNA1.1 · Official weights: Zyphra/ZUNA1.1 · Weights copies: modulsx/ZUNA1.1, suryatmodulus/ZUNA1.1, oraculumai/ZUNA1.1-CoreML-Apple, oraculumai/ZUNA1.1-CoreML-Apple-1s, oraculumai/ZUNA1.1-CoreML-Apple-30s, oraculumai/ZUNA1.1-Classifier-CoreML-Apple | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `Apache-2.0` (this site’s rights record) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | Pretraining exposure: unknown [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-zuna) | | **Other models and papers we track** — Other models and papers we track, each at the furthest step checked; a row that is held says why. | | | | | | AXON — [Code ↗](https://github.com/braindecode/braindecode) · [Weights ↗](https://huggingface.co/models?search=axon-eeg) — Official weights: MannasAI/axon-eeg | Foundation model · human scalp EEG | Code: `Apache-2.0, BSD-3-Clause` (the module’s file terms and the project’s root licence, in that order) Weights: `Apache-2.0` (Hugging Face licence tag) Data: not found | Model ladder · step 4 of 5 · Scored here — gap at step 2: data terms | Pretraining exposure: unknown; Its results on this site: [Pilot studies](https://bci.report/topics/pilot-studies/#eegmat-ten-second) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-axon) | | BandVQ — [Paper ↗](https://arxiv.org/html/2605.24921v1) | Foundation model · human scalp EEG | Code: not found Weights: not found Data: per source (the paper) | Model ladder · step 1 of 5 · Identified | Held: no official release of code or weights was found. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-bandvq) | | [BENDR](https://bci.report/methods/#bendr) — [Weights ↗](https://huggingface.co/models?search=braindecode-bendr) | Foundation model · human scalp EEG | Code: not found Weights: `BSD-3-Clause` (the model card’s metadata) Data: not found | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: code licence — gap at step 2: data terms | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-bendr) | | [BIOT](https://bci.report/methods/#biot) — [Code ↗](https://github.com/ycq091044/BIOT) | Foundation model · human scalp EEG | Code: `MIT` (this site’s rights record) Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-biot) | | BioX-Bridge — [ICLR 2026 paper ↗](https://proceedings.iclr.cc/paper_files/paper/2026/hash/b4ec8a0ae0373d694dc8529ca523e9a9-Abstract-Conference.html) · status checked 16 October 2026 — [Code ↗](https://github.com/chenqi-li/BioX-Bridge) | Adaptation method · human scalp EEG, ECG, PPG, and EMG | Code: not found Weights: per source (the paper) Data: per source (the paper) | Model ladder · step 1 of 5 · Identified | Held: no licence found for the code.; Held: no checkpoint is released. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-biox-bridge) | | Brain-OF — [Paper ↗](https://arxiv.org/abs/2602.23410) · [Code ↗](https://github.com/HanningGuo/BrainOF) · [Weights ↗](https://huggingface.co/models?search=Brain-OF-Base) — Also: Brain-OF Base, Brain-OF Large, BrainOF · Official weights: han-ning/Brain-OF-Base, han-ning/Brain-OF-Large | Foundation model · human scalp EEG, human fMRI, and human MEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `CC BY-NC 4.0` (Hugging Face licence tag) Data: per source (the paper) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-brain-of) | | BrainLM — [ICLR 2024 paper ↗](https://proceedings.iclr.cc/paper_files/paper/2024/hash/029ce70401321de3808b3ac39e1ab167-Abstract-Conference.html) · status checked 16 October 2026 — [Code ↗](https://github.com/vandijklab/BrainLM) · [Weights ↗](https://huggingface.co/models?search=brainlm) — Official weights: vandijklab/brainlm | Foundation model · human fMRI | Code: not found Weights: `CC BY-NC-ND 4.0` (the model card’s metadata) Data: on application (the paper) | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: code licence | Held: the weights’ licence is non-commercial with no derivatives.; Held: trained on signals other than human scalp EEG. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-brainlm) | | BrainPro — [Paper ↗](https://arxiv.org/abs/2509.22050) · [Code ↗](https://github.com/yi-ding-cs/BrainPro) · [Weights ↗](https://huggingface.co/models?search=BrainPro-V1) — Also: BrainPro-base, BrainPro-V1 · Official weights: dycs/BrainPro-V1 | Foundation model · human scalp EEG | Code: `CBCR License 1.0` (the repository’s licence file) Weights: `CBCR License 1.0` (the repository’s licence file) Data: per source (the paper) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-brainpro) | | [CSBrain](https://bci.report/methods/#csbrain) — [Code ↗](https://github.com/yuchen2199/CSBrain) | Foundation model · human scalp EEG | Code: not found Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-csbrain) | | DIVER-1 — [Paper ↗](https://arxiv.org/abs/2512.19097) · [Code ↗](https://github.com/DIVER-Project/DIVER-1) — Weights copies: braindecode/DIVER-1-0.1s-tiny, braindecode/DIVER-1-1s-small | Foundation model · human iEEG and human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | Held: no licence found for the weights. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-diver-1) | | EEG-Captioner — [Code ↗](https://github.com/lyJinAoki/EEG-Captioner) · [Weights ↗](https://huggingface.co/models?search=EEG-Captioner) — Official weights: LyJinAoki/EEG-Captioner | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `Apache-2.0` (Hugging Face licence tag) Data: not found | Model ladder · step 4 of 5 · Scored here — gap at step 2: data terms | Pretraining exposure: unknown; Its results on this site: [Pilot studies](https://bci.report/topics/pilot-studies/#eegmat-ten-second) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eeg-captioner) | | [EEG-Conformer](https://bci.report/methods/#eeg-conformer) — [Code ↗](https://github.com/eeyhsong/EEG-Conformer) | Trained from scratch · human scalp EEG | Code: `GPL-3.0` (this site’s rights record) Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eeg-conformer) | | [EEG-DINO](https://bci.report/methods/#eeg-dino) — [Paper ↗](https://papers.miccai.org/miccai-2025/0278-Paper3347.html) | Foundation model · human scalp EEG | Code: not found Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eeg-dino) | | EEG-PRIME — [Paper ↗](https://arxiv.org/html/2608.13072v1) | Foundation model · human scalp EEG | Code: not found Weights: not found Data: per source (the paper) | Model ladder · step 1 of 5 · Identified | Held: the code repository could not be reached. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eeg-prime) | | [EEGSimpleConv](https://bci.report/methods/#eegsimpleconv) — [Paper ↗](https://arxiv.org/abs/2309.07159) · [Code ↗](https://braindecode.org/stable/generated/braindecode.models.EEGSimpleConv.html) | Trained from scratch · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source (this site’s rights record) | Model ladder · step 4 of 5 · Scored here | Its results on this site: [Motor imagery & rest](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv), [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#signal-jepa-eegsimpleconv) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-eegsimpleconv) | | FAME — [Paper ↗](https://arxiv.org/html/2608.01898v1) — Also: Frequency-balanced Autoencoding for Masked EEG | Foundation model · human scalp EEG | Code: not found Weights: not found Data: per source (the paper) | Model ladder · step 1 of 5 · Identified | Held: no official release of code or weights was found. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-fame) | | LaBraM at Scale — [Paper ↗](https://arxiv.org/abs/2611.03165) · [Code ↗](https://github.com/AlbertMentat/LaBraM-at-Scale) · [Weights ↗](https://huggingface.co/models?search=LaBraM-at-Scale) — Also: LaBraM at Scale Quark-B, LaBraM at Scale Quark-L, LaBraM at Scale Quark-H, LaBraM at Scale Atom-L, LaBraM at Scale Molecule-L, LaBraM at Scale Quark-G · Official weights: Pan-Junyu/LaBraM-at-Scale | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (the repository’s licence file) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | Its results on this site: [Does pretraining help?](https://bci.report/topics/does-pretraining-help/#more-checkpoints), [LaBraM](https://bci.report/methods/labram/#labram-at-scale) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-labram-at-scale) | | LEAF — [Paper ↗](https://arxiv.org/abs/2509.24302) · [Code ↗](https://github.com/Jiang-Muyun/LEAF) · [Weights ↗](https://github.com/Jiang-Muyun/LEAF/releases/tag/v1.0) — Also: Language-EEG Aligned Foundation Model | Foundation model · human scalp EEG | Code: `CBCR License 1.0` (the repository’s licence file) Weights: `CBCR License 1.0` (the repository’s licence file) Data: not found | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: data terms | Pretraining exposure: label-level on Arithmetic & rest (its instruction-tuned checkpoint uses EEGMAT’s labels) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-leaf) | | LEGEND — [Code ↗](https://github.com/5GYYYYY/LEGEND) · [Weights ↗](https://huggingface.co/models?search=LEGEND) — Official weights: 5gyyyyy/LEGEND | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (Hugging Face licence tag) Data: not found | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: data terms | Pretraining exposure: unknown [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-legend) | | LuMamba — [Paper ↗](https://arxiv.org/abs/2603.19100) · [Code ↗](https://github.com/pulp-bio/BioFoundation) · [Weights ↗](https://huggingface.co/models?search=LuMamba) — Also: Latent Unified Mamba · Official weights: PulpBio/LuMamba | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `CC BY-ND 4.0` (Hugging Face licence tag) Data: per source (the paper) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-lumamba) | | MANAS-2 — [Paper ↗](https://arxiv.org/abs/2609.13717) · [Weights ↗](https://huggingface.co/models?search=MANAS-2) — Official weights: MannasAI/MANAS-2 | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `Apache-2.0` (Hugging Face licence tag) Data: per source (the paper) | Model ladder · step 2 of 5 · Terms recorded | Pretraining exposure: unknown [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-manas-2) | | MENDR — [Code ↗](https://github.com/MatthewChen37/MENDR-Manifold-Embedded-Neural-Data-Representations) · [Weights ↗](https://huggingface.co/models?search=MENDR-Manifold-Embedded-Neural-Data-Representations) — Also: Manifold-Embedded Neural Data Representations · Official weights: MatthewChen37/MENDR-Manifold-Embedded-Neural-Data-Representations | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (Hugging Face licence tag) Data: per source (the project page) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-mendr) | | [MIRepNet](https://bci.report/methods/#mirepnet) — [Code ↗](https://github.com/staraink/MIRepNet) | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `MIT` (the model card’s metadata) Data: not found | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: data terms | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-mirepnet) | | [Neuro-GPT](https://bci.report/methods/#neuro-gpt) — [Code ↗](https://github.com/wenhui0206/NeuroGPT) | Foundation model · human scalp EEG | Code: `GPL-3.0` (the repository’s licence file) Weights: `GPL-3.0` (the model card’s metadata) Data: not found | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: data terms | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-neuro-gpt) | | Neuroformer — ICLR 2024 paper · status checked 16 October 2026 — [Code ↗](https://github.com/a-antoniades/Neuroformer) | Generative model · mouse spiking | Code: `MIT` (the repository’s licence file) Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | Held: no checkpoint is released.; Held: trained on signals other than human scalp EEG. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-neuroformer) | | NeuroLM-B — [ICLR 2025 paper ↗](https://proceedings.iclr.cc/paper_files/paper/2025/file/8b4add8b0aa8749d80a34ca5d941c355-Paper-Conference.pdf) · status checked 16 October 2026 — [Code ↗](https://github.com/935963004/NeuroLM) · [Weights ↗](https://huggingface.co/models?search=NeuroLM) — Official weights: Weibang/NeuroLM | Foundation model · human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: `CC BY 4.0` (the model card’s metadata) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here | Pretraining exposure: unknown; Its results on this site: [Pilot studies](https://bci.report/topics/pilot-studies/#neurolm-eegmat), [Clinical groups](https://bci.report/topics/clinical-groups/#iesseeg) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-neurolm-b) | | NeuroRVQ — [Paper ↗](https://arxiv.org/abs/2510.13068) · [Code ↗](https://github.com/KonstantinosBarmpas/NeuroRVQ) · [Weights ↗](https://huggingface.co/models?search=NeuroRVQ) — Official weights: ntinosbarmpas/NeuroRVQ | Foundation model · human scalp EEG | Code: `CC BY-NC 4.0` (the repository’s licence file) Weights: `CC BY-NC 4.0` (the repository’s licence file) Data: per source (the paper) | Model ladder · step 4 of 5 · Scored here — gap at step 2: data terms | Pretraining exposure: unknown; Its results on this site: [Pilot studies](https://bci.report/topics/pilot-studies/#neurorvq) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-neurorvq) | | NOMAD-EEG-Large — [Code ↗](https://github.com/AbdulkaderHelwan/DecisionEEG-1) · [Weights ↗](https://github.com/AbdulkaderHelwan/DecisionEEG-1/releases/tag/v1.0) — Also: NOMAD-EEG, DecisionEEG-1 | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `Apache-2.0` (the repository’s licence file) Data: per source (the project page) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-nomad-eeg-large) | | NSP — [Paper ↗](https://arxiv.org/abs/2609.31167v1) — Also: Neural State Prediction | Foundation model · human scalp EEG | Code: `MIT (reference code, not linked)` (this site’s rights record) Weights: not found Data: per source (the paper) | Model ladder · step 3 of 5 · Files checked — gap at step 2: weights licence — gap at step 2: file listing | Held: no checkpoint was found in the sources checked.; Its results on this site: [Pilot studies](https://bci.report/topics/pilot-studies/#nsp-components) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-nsp) | | OmniMouse — ICLR 2026 paper · status checked 16 October 2026 — [Code ↗](https://github.com/enigma-brain/omnimouse) · [Weights ↗](https://huggingface.co/models?search=omnimouse-80M) · [Data ↗](https://huggingface.co/datasets?search=omnimouse-dataset) — Official weights: the-enigma-project/omnimouse-80M · Weights copies: the-enigma-project/omnimouse, TheKonstantinWilleke/omnimouse | Foundation model · mouse calcium imaging | Code: not found Weights: `Apache-2.0` (the model card’s metadata) Data: `CC BY 4.0` (the dataset card) | Model ladder · step 2 of 5 · Terms recorded — gap at step 2: code licence | Held: no licence found for the code.; Held: trained on signals other than human scalp EEG. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-omnimouse) | | PanLUNA — [Paper ↗](https://arxiv.org/abs/2604.04297) · [Code ↗](https://github.com/pulp-bio/BioFoundation) · [Weights ↗](https://huggingface.co/models?search=PanLUNA) — Official weights: PulpBio/PanLUNA | Foundation model · human scalp EEG, ECG, and PPG | Code: `Apache-2.0` (the repository’s licence file) Weights: `CC BY-ND 4.0` (Hugging Face licence tag) Data: per source (the paper) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-panluna) | | Population Transformer (PopT) — ICLR 2025 paper · status checked 16 October 2026 — [Code ↗](https://github.com/czlwang/PopulationTransformer) · [Weights ↗](https://huggingface.co/models?search=popt_brainbert_stft) · [Data ↗](https://braintreebank.dev/) — Official weights: PopulationTransformer/popt_brainbert_stft · Weights copies: braindecode/popt-pretrained | Spatial aggregation model · human iEEG and human scalp EEG | Code: `MIT` (the repository’s licence file) Weights: not found Data: `CC BY 4.0` (the project page) | Model ladder · step 1 of 5 · Identified | Held: no licence found for the weights. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-popt) | | POYO-CAP — [ICLR 2026 paper ↗](https://proceedings.iclr.cc/paper_files/paper/2026/hash/e6d82f5b06481dabbdf74bd38d013bbe-Abstract-Conference.html) · status checked 16 October 2026 — [Code ↗](https://github.com/Transconnectome/POYO-CAP) · [Data ↗](https://portal.brain-map.org/explore/circuits/observatory) | Decoding model · mouse calcium imaging | Code: `MIT` (the repository’s licence file) Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | Held: no checkpoint is released.; Held: trained on signals other than human scalp EEG. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-poyo-cap) | | POYO+ — [ICLR 2025 paper ↗](https://proceedings.iclr.cc/paper_files/paper/2025/hash/953390c834451505703c9da45de634d8-Abstract-Conference.html) · status checked 16 October 2026 — [Data ↗](https://portal.brain-map.org/explore/circuits/observatory) | Decoding model · mouse calcium imaging | Code: not found Weights: not found Data: not found | Model ladder · step 1 of 5 · Identified | Held: the code repository could not be reached.; Held: trained on signals other than human scalp EEG. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-poyo-plus) | | S-CEReBrO — [Paper ↗](https://arxiv.org/abs/2607.27913) · [Code ↗](https://github.com/pulp-bio/BioFoundation) · [Weights ↗](https://huggingface.co/models?search=S-CEReBrO) — Also: CEReBrO · Official weights: PulpBio/S-CEReBrO | Foundation model · human scalp EEG | Code: `Apache-2.0` (the repository’s licence file) Weights: `CC BY-ND 4.0` (Hugging Face licence tag) Data: per source (the paper) | Model ladder · step 2 of 5 · Terms recorded | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-s-cerebro) | | [Signal-JEPA](https://bci.report/methods/#signal-jepa) — [Paper ↗](https://arxiv.org/abs/2403.11772) · [Weights ↗](https://huggingface.co/models?search=signal-jepa) | Foundation model · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: `MIT` (the model card’s metadata) Data: not found | Model ladder · step 4 of 5 · Scored here — gap at step 2: data terms | Its results on this site: [Motor imagery & rest](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv), [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#signal-jepa-eegsimpleconv) [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-signal-jepa) | | [Spectral ridge](https://bci.report/methods/#other-evaluated) — [Code ↗](https://scikit-learn.org/stable/) | Classical method · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-spectral-ridge) | | STST-JEPA — [Paper ↗](https://arxiv.org/abs/2607.06629) — Also: Shallow-Target Spatio-Temporal JEPA | Foundation model · human scalp EEG | Code: not found Weights: not found Data: per source (the paper) | Model ladder · step 1 of 5 · Identified | Held: no official release of code or weights was found. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-stst-jepa) | | [Temporal ridge](https://bci.report/methods/#other-evaluated) — [Code ↗](https://scikit-learn.org/stable/) | Classical method · human scalp EEG | Code: `BSD-3-Clause` (this site’s rights record) Weights: not applicable (this site’s rights record) Data: per source | Model ladder · step 5 of 5 · In the matrix | [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-temporal-ridge) | | USE-FM — [Paper ↗](https://arxiv.org/html/2610.05503v1) — Also: Universal Semantic EEG Foundation Model | Foundation model · human scalp EEG | Code: not found Weights: not found Data: per source (the paper) | Model ladder · step 1 of 5 · Identified | Held: no official release of code or weights was found. [Correct this row](mailto:contact@bci.report?subject=Register%20row%20r-model-use-fm) | Download: [`open-register-20261106.json`](https://bci.report/data/open-register-20261106.json) (both ladders, with each step’s evidence and the findings as codes). [Every file](https://bci.report/api/) ## Cite this page BCI Report (2026). *Every EEG model and paper we track, and how far we checked it*. https://bci.report/methods/register/ Figures from release [`open-register-update-20261106`](https://bci.report/releases/#open-register-update-20261106) (2026-11-06). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/register/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # REVE: results on public EEG datasets Foundation model REVE is a pretrained EEG encoder released in Small, Base and Large sizes, all trained on one corpus that its paper's appendix lists: public collections from TUH, PhysioNet, OpenNeuro and MOABB among others, with the recordings of the authors' own downstream tasks removed. BCI Report ran the Base and Large checkpoints frozen, with the published ridge head, and the Small encoder, an early checkpoint (epoch 15 of 200), in the same way; their input removes each segment's per-channel mean, a deviation declared before scoring. The weights are under the REVE Responsible Use License, accepted for this site on 4 October 2026. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** REVE Base · REVE Large · brain-bzh/reve-base · brain-bzh/reve-large · REVE Small · brain-bzh/reve-small **Reference** [REVE, arXiv:2510.21585 ↗](https://arxiv.org/abs/2510.21585) Directory status: Evaluated · [Official code ↗](https://huggingface.co/brain-bzh/reve-base) · Description sources [arxiv.org](https://arxiv.org/abs/2510.21585) · [huggingface.co](https://huggingface.co/datasets/brain-bzh/reve-dataset) · [huggingface.co](https://huggingface.co/brain-bzh/reve-base) · [huggingface.co](https://huggingface.co/brain-bzh/reve-large) · [huggingface.co](https://huggingface.co/brain-bzh/reve-small) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 66.2% (61.8%–70.6%) - REVE Large: 70.2% (64.0%–76.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 66.2% (61.8%–70.6%) | 10 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 70.2% (64.0%–76.1%) | 10 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 68.8% (64.1%–73.2%) - REVE Large: 69.0% (63.3%–74.7%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 68.8% (64.1%–73.2%) | 36 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 69.0% (63.3%–74.7%) | 36 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 4 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base · Frozen encoder + trained head: 64.0% (60.4%–67.7%) - REVE Base · LoRA rank 4 + head: 66.3% (61.9%–70.7%) - REVE Large · Frozen encoder + trained head: 65.6% (61.1%–70.1%) - REVE Large · LoRA rank 4 + head: 69.2% (63.4%–74.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + trained head | Balanced accuracy | 64.0% (60.4%–67.7%) — 1,026 trainable parameters. | 36 | | [REVE Base](https://bci.report/methods/reve/) | LoRA rank 4 + head | Balanced accuracy | 66.3% (61.9%–70.7%) — 181,250 trainable parameters. | 36 | | [REVE Base](https://bci.report/methods/reve/) | LoRA minus frozen + head, same people and folds | Paired difference | +2.3 pp (+0.1 pp–+4.6 pp) — 21 people improved, 13 got worse, 2 unchanged. The interval excludes zero. | 36 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + trained head | Balanced accuracy | 65.6% (61.1%–70.1%) — 2,434 trainable parameters. | 36 | | [REVE Large](https://bci.report/methods/reve/) | LoRA rank 4 + head | Balanced accuracy | 69.2% (63.4%–74.9%) — 430,466 trainable parameters. | 36 | | [REVE Large](https://bci.report/methods/reve/) | LoRA minus frozen + head, same people and folds | Paired difference | +3.6 pp (+0.5 pp–+6.8 pp) — 21 people improved, 13 got worse, 2 unchanged. The interval excludes zero. | 36 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 37.5% (33.3%–41.9%) - REVE Large: 52.7% (47.7%–57.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 37.5% (33.3%–41.9%) | 70 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 52.7% (47.7%–57.8%) | 70 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 25.5% (22.4%–28.8%) - REVE Large: 46.4% (41.7%–51.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 25.5% (22.4%–28.8%) | 70 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 46.4% (41.7%–51.2%) | 70 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 50.5% (48.5%–52.3%) - REVE Large: 54.6% (52.7%–56.5%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 50.5% (48.5%–52.3%) | 21 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 54.6% (52.7%–56.5%) | 21 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 53.7% (51.8%–55.4%) - REVE Large: 59.7% (57.5%–61.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 53.7% (51.8%–55.4%) | 30 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 59.7% (57.5%–61.8%) | 30 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - REVE Base: 74.6% (72.9%–76.4%) - REVE Large: 76.6% (75.2%–78.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 74.6% (72.9%–76.4%) | 20 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + ridge head | Balanced accuracy | 76.6% (75.2%–78.0%) | 20 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + linear head | Commands detected ≤3 s | 12 / 60 — 0 of 4 people always abstained. | 4 | | [REVE Base](https://bci.report/methods/reve/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + linear head | Commands detected ≤3 s | 6 / 60 — 2 of 4 people always abstained. | 4 | | [REVE Large](https://bci.report/methods/reve/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | Weights terms: REVE under the REVE Responsible Use License v1.0 (model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7). [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### REVE Base - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 69.19M - Revision: `brain-bzh/reve-base @ dc2a075c` - Checkpoint SHA-256: `8ecc650619598748286c2457f81f5c6bd12e8bb59db44f7b02af1955c44de8fe` - Paper: [paper ↗](https://arxiv.org/abs/2510.21585) - Weights licence: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) - Licence note: REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. - Rights review: Accepted by the owner on 2026-10-04. Aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; the REVE paper is cited and the model version named; no adapted weights are shared. - Row footnote: Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Notes: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ### REVE Large - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 389.98M - Revision: `brain-bzh/reve-large @ 317531c7` - Checkpoint SHA-256: `f519428404aab27f7db816722c7fb8c73d08daf10986c9c21b0066be9e93bb31` - Paper: [paper ↗](https://arxiv.org/abs/2510.21585) - Weights licence: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) - Licence note: REVE Responsible Use License v1.0 (as REVE Base). - Rights review: Accepted by the owner on 2026-10-04. Aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; the REVE paper is cited and the model version named; no adapted weights are shared. - Row footnote: Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Notes: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. ### REVE Base · REVE Large | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) · [source ↗](https://github.com/NeuroTechX/moabb/commits/develop/moabb/datasets/ssvep_liu2020.py) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2510.21585) · [source ↗](https://huggingface.co/datasets/brain-bzh/reve-dataset) · [source ↗](https://github.com/NeuroTechX/moabb/commits/develop/moabb/datasets/triana_guzman2024.py) | **Note on the check:** REVE's Table 7 counts 27 MOABB datasets but names 25 (26 with the open-subset card); the unnamed slot cannot be BETA or ds005342 because MOABB added those loaders only in March 2026. ## REVE Small, an early checkpoint REVE Small's encoder was released at epoch 15 of a 200-epoch schedule; every REVE Small value carries the label "early checkpoint (epoch 15 of 200)". ### REVE Small - Parameters: 12.67M - Pooled width: 512 - Checkpoint: `brain-bzh/reve-small @ 3565bdcb` · `model.safetensors` - SHA-256: `e573ec80cc21f351e60d899920d112e1550c01930eb6b958d6bc7ff1d0ef2b7a` - Input: each segment's per-channel mean removed, then divided by 100 - Pooling: attention pooling with the pretrained query token - Positions: `brain-bzh/reve-positions @ befa5b57` - Backend: Local Ubuntu / CPU - Label: early checkpoint (epoch 15 of 200) - Input: REVE Small reads its input as the REVE Base and Large rows do: each segment's per-channel mean is removed before dividing by 100; no sensitivity run was made for Small. - Exposure: REVE Small's card does not restate its corpus; its exposure is read from the REVE corpus list (checked 2026-10-04), which includes none of the core datasets. Its cells are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-reve-small)” and on each protocol page. Terms: the REVE Responsible Use License v1.0 (model version reve-small @ 3565bdcb), aggregate scientific results only; cite [arXiv 2510.21585](https://arxiv.org/abs/2510.21585). No weights are distributed. [foundation-checkpoints-update.json](https://bci.report/data/foundation-checkpoints-update.json) · [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Cite this page BCI Report (2026). *REVE: results on public EEG datasets*. https://bci.report/methods/reve/ Figures from releases [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/reve/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # TS+LR and MDM: results on public EEG datasets Classical method Riemannian classification works on the spatial covariance matrix of each EEG window, a point on the manifold of symmetric positive-definite matrices. Two decoders from Barachant, Bonnet, Congedo and Jutten (IEEE Transactions on Biomedical Engineering, 2012) are run here: the minimum distance to mean (MDM) classifier, which assigns a window to the class whose mean covariance is nearest in Riemannian distance, and tangent-space logistic regression (TS+LR), which maps covariances into the tangent space at a reference point and fits a logistic regression there. BCI Report runs both through pyRiemann with fixed settings and OAS covariance estimates, under the same participant-held-out protocols as the other methods; the settings follow the MOABB reference pipelines (TSLR.yml with max_iter raised, MDM.yml unchanged). Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** TS+LR · tangent space + logistic regression · tangent-space logistic regression · MDM · minimum distance to mean · Barachant et al. 2012 **Reference** [Barachant A, Bonnet S, Congedo M, Jutten C. Multiclass brain–computer interface classification by Riemannian geometry. IEEE Transactions on Biomedical Engineering 59(4):920–928, 2012. ↗](https://doi.org/10.1109/TBME.2011.2172210) [Official code ↗](https://github.com/pyRiemann/pyRiemann) · Description sources [doi.org](https://doi.org/10.1109/TBME.2011.2172210) · [github.com](https://github.com/pyRiemann/pyRiemann) · [github.com](https://github.com/NeuroTechX/moabb) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [core-matrix-20261106-mi-rest.csv](https://bci.report/data/core-matrix-20261106-mi-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - TS+LR · OAS · 4–40 Hz: 58.9% (54.9%–63.9%) - MDM · OAS · 4–40 Hz: 53.8% (50.8%–57.9%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [TS+LR · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Rest versus right-hand imagery | Balanced accuracy | 58.9% (54.9%–63.9%) — Primary row · one fixed configuration (MOABB-reference settings) | 10 | | [MDM · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Rest versus right-hand imagery | Balanced accuracy | 53.8% (50.8%–57.9%) — Secondary row · one fixed configuration (MOABB-reference settings) | 10 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - TS+LR · OAS · 4–40 Hz: 65.5% (59.6%–71.8%) - MDM · OAS · 4–40 Hz: 71.5% (65.5%–77.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [TS+LR · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Serial subtraction versus resting EEG | Balanced accuracy | 65.5% (59.6%–71.8%) — Primary row · one fixed configuration (MOABB-reference settings) | 36 | | [MDM · OAS · 4–40 Hz](https://bci.report/methods/riemannian/) | Serial subtraction versus resting EEG | Balanced accuracy | 71.5% (65.5%–77.8%) — Secondary row · one fixed configuration (MOABB-reference settings) | 36 | ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [core-matrix-20261106-idle.csv](https://bci.report/data/core-matrix-20261106-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [TS+LR · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 35.0% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | | [TS+LR · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Idle false activation | 6.7% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | | [MDM · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 35.0% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | | [MDM · OAS](https://bci.report/methods/riemannian/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% — Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. | 4 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) Core matrix ## Core-matrix rows: recipe and paired differences Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) ### How these rows were run | Method | Recipe | MOABB reference pipeline | Implementation | Code licence | | --- | --- | --- | --- | --- | | TS+LR — Supervised fit | OAS · 4–40 Hz; OAS · idle-native windows | MOABB TSLR.yml, with max_iter raised to 5000 | pyRiemann 0.12 (wheel SHA-256 74a90b2ef0938c94fe9bf90b6506991fc19da55850aae5fdf992026b3845d43a) and scikit-learn | BSD-3-Clause | | MDM — Supervised fit | OAS · 4–40 Hz; OAS · idle-native windows | MOABB MDM.yml, unchanged | pyRiemann 0.12 (wheel SHA-256 74a90b2ef0938c94fe9bf90b6506991fc19da55850aae5fdf992026b3845d43a) and scikit-learn | BSD-3-Clause | One fixed configuration per method: library or MOABB-reference settings with the stated deviations. The 4–40 Hz band-pass (or demeaning plus 50 Hz decimation on the ERP protocols) is part of the recipe and is applied to each window on its own. The EEGMAT signals were ICA-cleaned and may be rank-deficient; OAS shrinkage then dominates the low-variance directions of the covariances that TS+LR and MDM read. ### Sensitivity lines The same methods with other input filtering, labelled; they are not matrix rows. - **TS+LR · OAS · raw** · [Motor imagery & rest](https://bci.report/protocols/mi-rest/): 54.4% (51.1%–57.9%), macro F1 0.473, 10 people - **TS+LR · OAS · 8–32 Hz** · [Motor imagery & rest](https://bci.report/protocols/mi-rest/): 60.1% (54.8%–65.6%), macro F1 0.530, 10 people - **MDM · OAS · raw** · [Motor imagery & rest](https://bci.report/protocols/mi-rest/): 52.0% (50.5%–53.7%), macro F1 0.438, 10 people - **MDM · OAS · 8–32 Hz** · [Motor imagery & rest](https://bci.report/protocols/mi-rest/): 56.0% (52.1%–60.6%), macro F1 0.460, 10 people - **TS+LR · OAS · raw** · [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/): 63.1% (57.5%–68.7%), macro F1 0.580, 36 people - **MDM · OAS · raw** · [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/): 70.7% (63.1%–78.2%), macro F1 0.667, 36 people ### Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. Paired participant bootstrap (10,000 draws). The four contrasts against the Spectral ridge cells carry 98.75 % intervals (Bonferroni over four); the other contrasts carry nominal 95 % intervals with no familywise claim. 'Higher' or 'lower' means the interval lies wholly above or below zero; 'unresolved' means it includes zero. Ten people: a percentile bootstrap over ten people tends to give intervals narrower than their nominal level. ### Primary contrasts: against the Spectral ridge cell | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | TS+LR − Spectral ridge — Motor imagery & rest — TS+LR · OAS · 4–40 Hz | +4.0 pp | −1.1 to 11.5 | −0.3 to 9.7 | unresolved | | TS+LR − Spectral ridge — Arithmetic & rest — TS+LR · OAS · 4–40 Hz | +8.8 pp | 1.6 to 16.1 | 3.0 to 14.6 | higher | ### Against EEGNet against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | TS+LR − EEGNet — Motor imagery & rest — TS+LR · OAS · 4–40 Hz | −12.1 pp | −18.1 to −5.4 | lower | | MDM − EEGNet — Motor imagery & rest — MDM · OAS · 4–40 Hz | −17.2 pp | −23.7 to −10.6 | lower | | TS+LR − EEGNet — Arithmetic & rest — TS+LR · OAS · 4–40 Hz | −2.1 pp | −8.0 to 3.8 | unresolved | | MDM − EEGNet — Arithmetic & rest — MDM · OAS · 4–40 Hz | +3.8 pp | −1.3 to 9.3 | unresolved | ### Secondary rows against the Spectral ridge cell | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | MDM − Spectral ridge — Motor imagery & rest — MDM · OAS · 4–40 Hz | −1.1 pp | −4.7 to 3.8 | unresolved | | MDM − Spectral ridge — Arithmetic & rest — MDM · OAS · 4–40 Hz | +14.7 pp | 8.9 to 20.5 | higher | ### Against the CSP+LDA row Show the table | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | TS+LR − CSP+LDA — Motor imagery & rest — TS+LR · OAS · 4–40 Hz | +11.8 pp | 7.9 to 16.8 | higher | | TS+LR − CSP+LDA — Motor imagery & rest — TS+LR · OAS · raw | +7.3 pp | 3.8 to 11.6 | higher | | TS+LR − CSP+LDA — Motor imagery & rest — TS+LR · OAS · 8–32 Hz | +13.0 pp | 7.8 to 18.6 | higher | | MDM − CSP+LDA — Motor imagery & rest — MDM · OAS · 4–40 Hz | +6.8 pp | 3.3 to 10.8 | higher | | MDM − CSP+LDA — Motor imagery & rest — MDM · OAS · raw | +4.9 pp | 1.3 to 9.2 | higher | | MDM − CSP+LDA — Motor imagery & rest — MDM · OAS · 8–32 Hz | +8.9 pp | 4.6 to 13.7 | higher | The CSP+LDA row reads the windows with no band-pass; the raw-arm lines read the same input. ### Against the frozen encoders LaBraM and CBraMod | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | TS+LR − LaBraM — Motor imagery & rest — TS+LR · OAS · 4–40 Hz | +5.4 pp | 0.3 to 11.3 | higher | | TS+LR − CBraMod — Motor imagery & rest — TS+LR · OAS · 4–40 Hz | −4.0 pp | −8.4 to 0.9 | unresolved | | TS+LR − LaBraM — Arithmetic & rest — TS+LR · OAS · 4–40 Hz | +0.9 pp | −4.6 to 6.6 | unresolved | | TS+LR − CBraMod — Arithmetic & rest — TS+LR · OAS · 4–40 Hz | +3.2 pp | −1.5 to 8.0 | unresolved | These pairs use the matrix's LaBraM and CBraMod rows (frozen encoder + ridge head) only. ### Recipe arms: the same method with other input filtering Show the table | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | TS+LR · OAS · 4–40 Hz − TS+LR · OAS · raw — Motor imagery & rest | +4.5 pp | 0.6 to 7.9 | higher | | MDM · OAS · 4–40 Hz − MDM · OAS · raw — Motor imagery & rest | +1.8 pp | −1.5 to 6.3 | unresolved | | TS+LR · OAS · 4–40 Hz − TS+LR · OAS · 8–32 Hz — Motor imagery & rest | −1.2 pp | −6.0 to 2.6 | unresolved | | MDM · OAS · 4–40 Hz − MDM · OAS · 8–32 Hz — Motor imagery & rest | −2.2 pp | −3.8 to −0.8 | lower | | TS+LR · OAS · 4–40 Hz − TS+LR · OAS · raw — Arithmetic & rest | +2.4 pp | −3.0 to 8.1 | unresolved | | MDM · OAS · 4–40 Hz − MDM · OAS · raw — Arithmetic & rest | +0.7 pp | −5.3 to 7.2 | unresolved | ### Within the family: primary row against secondary row | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | TS+LR − MDM — Motor imagery & rest — TS+LR · OAS · 4–40 Hz | +5.1 pp | 2.3 to 7.8 | higher | | TS+LR − MDM — Arithmetic & rest — TS+LR · OAS · 4–40 Hz | −6.0 pp | −11.6 to −0.5 | lower | ### Credits TS+LR and MDM: Barachant, Bonnet, Congedo and Jutten, IEEE TBME 2012 and Neurocomputing 2013; xDAWN: Rivet, Souloumiac, Attina and Gibert, IEEE TBME 2009; XDAWNCov: Barachant and Congedo 2014; OAS: Chen, Wiesel, Eldar and Hero, IEEE TSP 2010; Ledoit and Wolf 2004; MOABB reference pipelines (BSD-3-Clause), Jayaram and Barachant, J. Neural Eng. 2018, read for citation and not executed; pyRiemann 0.12 (BSD-3-Clause). - Barachant, Bonnet, Congedo and Jutten. Multiclass brain–computer interface classification by Riemannian geometry. IEEE Transactions on Biomedical Engineering 59(4), 2012. [doi ↗](https://doi.org/10.1109/TBME.2011.2172210) - Barachant, Bonnet, Congedo and Jutten. Classification of covariance matrices using a Riemannian-based kernel for BCI applications. Neurocomputing 112, 2013. [doi ↗](https://doi.org/10.1016/j.neucom.2012.12.039) - Chen, Wiesel, Eldar and Hero. Shrinkage algorithms for MMSE covariance estimation. IEEE Transactions on Signal Processing 58(10), 2010. [doi ↗](https://doi.org/10.1109/TSP.2010.2053029) - Jayaram and Barachant. MOABB: trustworthy algorithm benchmarking for BCIs. Journal of Neural Engineering 15(6), 2018. [doi ↗](https://doi.org/10.1088/1741-2552/aadea0) ## Cite this page BCI Report (2026). *TS+LR and MDM: results on public EEG datasets*. https://bci.report/methods/riemannian/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/riemannian/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ShallowFBCSPNet: results on public EEG datasets Compact model ShallowFBCSPNet is Braindecode's name for the shallow ConvNet of Schirrmeister et al. (Human Brain Mapping, 2017). Modeled on filter bank common spatial patterns (FBCSP), it applies a temporal convolution, a spatial convolution across electrodes, squaring, mean pooling and a logarithm, which together approximate log band-power features. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** ShallowFBCSPNet · Shallow ConvNet · ShallowConvNet · Schirrmeister et al. 2017 **Reference** [Schirrmeister RT, Springenberg JT, Fiederer LDJ, Glasstetter M, Eggensperger K, Tangermann M, Hutter F, Burgard W, Ball T. Deep learning with convolutional neural networks for EEG decoding and visualization. Human Brain Mapping 38(11):5391–5420, 2017. ↗](https://doi.org/10.1002/hbm.23730) Directory status: Evaluated · [Official code ↗](https://github.com/braindecode/braindecode/tree/v1.5.1) · Description sources [arxiv.org](https://arxiv.org/abs/1703.05051) · [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC5655781/) · [doi.org](https://doi.org/10.1002/hbm.23730) · [github.com](https://github.com/braindecode/braindecode/blob/v1.5.1/braindecode/models/shallow_fbcsp.py) · [github.com](https://github.com/robintibor/braindecode) · [github.com](https://github.com/TNTLFreiburg/braindecode) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Core matrix · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/) · Chance level 50.0% · [core-matrix-20261106-mi-rest.csv](https://bci.report/data/core-matrix-20261106-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Rest versus right-hand imagery | Balanced accuracy | 68.8% (63.4%–73.9%) — One fixed seed and training budget, as the EEGNet cell | 10 | ### [EEGMAT](https://bci.report/datasets/eegmat/) · Core matrix · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) · Chance level 50.0% · [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Serial subtraction versus resting EEG | Balanced accuracy | 67.9% (63.4%–72.4%) — One fixed seed and training budget, as the EEGNet cell | 36 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · Chance level 2.5% · [core-matrix-20261106-beta-8ch.csv](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 23.0% (20.3%–25.9%) — One fixed seed and training budget, as the EEGNet cell | 70 | ### [BETA](https://bci.report/datasets/beta/) · Core matrix · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) · Chance level 2.5% · [core-matrix-20261106-beta-4ch.csv](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 16.2% (14.1%–18.6%) — One fixed seed and training budget, as the EEGNet cell | 70 | ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Visual target versus nontarget events | Balanced accuracy | 49.7% (49.0%–50.4%) — One fixed seed and training budget, as the EEGNet cell | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Reading · target versus nontarget events | Balanced accuracy | 56.7% (54.7%–58.8%) — One fixed seed and training budget, as the EEGNet cell | 30 | ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Core matrix · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/) · Chance level 20.0% · [core-matrix-20261106-sleep-scalp.csv](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Five stages · balanced scalp-EEG sample | Balanced accuracy | 55.0% (53.1%–56.8%) — One fixed seed and training budget, as the EEGNet cell | 20 | ### [ds005342](https://bci.report/datasets/ds005342/) · Core matrix · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/) · [experiments.json](https://bci.report/data/experiments.json) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Seated motor imagery · cue-gated replay | Command detection ≤3 s | 43.3% | 4 | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) | Seated motor imagery · cue-gated replay | Idle false activation | 1.7% | 4 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Cite this page BCI Report (2026). *ShallowFBCSPNet: results on public EEG datasets*. https://bci.report/methods/shallowfbcspnet/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/shallowfbcspnet/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # SingLEM: results on public EEG datasets Foundation model SingLEM is a single-channel EEG foundation model: a tokenizer encodes each channel on its own. Its paper lists 71 public pretraining datasets, TMNRED among them, so its cells on the semantic-target protocol are marked as in the authors' pretraining list. BCI Report ran the primary checkpoint, the one that excludes the authors' own downstream sources, frozen, and adapted it on EEGMAT. Its tokenizer does not see the last quarter-second of each two-second segment. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** ttlabtuat/SingLEM **Reference** [SingLEM, arXiv:2509.17920 ↗](https://arxiv.org/abs/2509.17920v3) Directory status: Evaluated · [Official code ↗](https://github.com/ttlabtuat/SingLEM) · Description sources [arxiv.org](https://arxiv.org/abs/2509.17920v3) · [github.com](https://github.com/ttlabtuat/SingLEM) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 72.0% (66.6%–76.6%) | 10 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 51.5% (49.2%–53.8%) | 36 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - SingLEM · Frozen encoder + trained head: 50.9% (49.5%–52.3%) - SingLEM · LoRA rank 4 + head: 60.6% (56.7%–64.4%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + trained head | Balanced accuracy | 50.9% (49.5%–52.3%) — 1,218 trainable parameters. | 36 | | [SingLEM](https://bci.report/methods/singlem/) | LoRA rank 4 + head | Balanced accuracy | 60.6% (56.7%–64.4%) — 33,986 trainable parameters. | 36 | | [SingLEM](https://bci.report/methods/singlem/) | LoRA minus frozen + head, same people and folds | Paired difference | +9.7 pp (+6.1 pp–+13.3 pp) — 30 people improved, 6 got worse, 0 unchanged. The interval excludes zero. | 36 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 26.6% (23.7%–29.5%) | 70 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 20.5% (17.8%–23.3%) | 70 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 51.2% (49.2%–53.1%) | 21 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 58.5% (56.1%–61.1%) — Pretraining exposure: in the authors' published pretraining list. | 30 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + ridge head | Balanced accuracy | 35.3% (33.7%–37.0%) | 20 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + linear head | Commands detected ≤3 s | 6 / 60 — 2 of 4 people always abstained. | 4 | | [SingLEM](https://bci.report/methods/singlem/) | Frozen encoder + linear head | Idle false activations | 5 / 60 | 4 | Weights terms: SingLEM under MIT. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### SingLEM - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 3.27M - Revision: `ttlabtuat/SingLEM @ 6db5b5af` - Checkpoint SHA-256: `cb4c1b4f4a6fae99a984b883177e09b8c2fe44ed0d4c9a0ded3928dc425584bb` - Paper: [paper ↗](https://arxiv.org/abs/2509.17920) - Weights licence: MIT (checkpoint committed in the MIT repository) - Licence note: MIT. - Rights review: Checkpoint committed in the MIT repository; aggregate scores only. - Row footnote: semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Notes: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2509.17920v3) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | in the authors' published pretraining list | [source ↗](https://arxiv.org/abs/2509.17920v3) · [source ↗](https://github.com/ttlabtuat/SingLEM) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2509.17920v3) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://arxiv.org/abs/2509.17920v3) | **Note on the check:** SingLEM and LaBraM lists are matched by dataset name and reference; most rows carry no accession number. ## Cite this page BCI Report (2026). *SingLEM: results on public EEG datasets*. https://bci.report/methods/singlem/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/singlem/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ST-EEGFormer: results on public EEG datasets Foundation model ST-EEGFormer is a pretrained EEG Transformer released in small, base and large sizes. Its paper's appendix lists eleven public pretraining datasets — BETA among them — plus the authors' own in-house recordings, so its cells on both BETA protocols are marked as in the authors' pretraining list. BCI Report ran Base and Large frozen and adapted Base on EEGMAT. Its README reserves the paper, the diagrams and the name, none of which is reproduced here. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** ST-EEGFormer Base · ST-EEGFormer Large · STEEGFormer **Reference** [ST-EEGFormer, OpenReview 5Xwm8e6vbh ↗](https://openreview.net/forum?id=5Xwm8e6vbh) Directory status: Evaluated · [Official code ↗](https://github.com/LiuyinYang1101/STEEGFormer) · Description sources [openreview.net](https://openreview.net/forum?id=5Xwm8e6vbh) · [github.com](https://github.com/LiuyinYang1101/STEEGFormer) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 58.1% (53.7%–63.3%) - ST-EEGFormer Large: 62.4% (57.4%–67.8%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 58.1% (53.7%–63.3%) | 10 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 62.4% (57.4%–67.8%) | 10 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 64.5% (59.0%–70.0%) - ST-EEGFormer Large: 63.4% (57.8%–69.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 64.5% (59.0%–70.0%) | 36 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 63.4% (57.8%–69.0%) | 36 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base · Frozen encoder + trained head: 63.2% (59.7%–66.7%) - ST-EEGFormer Base · LoRA rank 4 + head: 65.1% (61.0%–69.4%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + trained head | Balanced accuracy | 63.2% (59.7%–66.7%) — 1,538 trainable parameters. | 36 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | LoRA rank 4 + head | Balanced accuracy | 65.1% (61.0%–69.4%) — 148,994 trainable parameters. | 36 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | LoRA minus frozen + head, same people and folds | Paired difference | +1.9 pp (−0.2 pp–+4.1 pp) — 20 people improved, 16 got worse, 0 unchanged. The interval includes zero: no change is established. | 36 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 17.1% (15.5%–18.8%) - ST-EEGFormer Large: 22.6% (20.1%–25.3%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 17.1% (15.5%–18.8%) — Pretraining exposure: in the authors' published pretraining list. | 70 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 22.6% (20.1%–25.3%) — Pretraining exposure: in the authors' published pretraining list. | 70 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 17.5% (15.8%–19.3%) - ST-EEGFormer Large: 22.4% (19.9%–25.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 17.5% (15.8%–19.3%) — Pretraining exposure: in the authors' published pretraining list. | 70 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 22.4% (19.9%–25.0%) — Pretraining exposure: in the authors' published pretraining list. | 70 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 53.6% (52.3%–54.8%) - ST-EEGFormer Large: 52.7% (51.3%–54.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 53.6% (52.3%–54.8%) | 21 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 52.7% (51.3%–54.2%) | 21 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 54.6% (52.9%–56.4%) - ST-EEGFormer Large: 57.8% (56.0%–59.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 54.6% (52.9%–56.4%) | 30 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 57.8% (56.0%–59.6%) | 30 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ST-EEGFormer Base: 77.3% (75.6%–78.8%) - ST-EEGFormer Large: 79.6% (78.2%–81.0%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 77.3% (75.6%–78.8%) | 20 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + ridge head | Balanced accuracy | 79.6% (78.2%–81.0%) | 20 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Commands detected ≤3 s | 16 / 60 — 1 of 4 people always abstained. | 4 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Commands detected ≤3 s | 20 / 60 — 1 of 4 people always abstained. | 4 | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) | Frozen encoder + linear head | Idle false activations | 1 / 60 | 4 | Weights terms: ST-EEGFormer under MIT. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### ST-EEGFormer Base - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 85.18M - Revision: `official GitHub release checkpoint-288.pth (code @ 542ee17)` - Checkpoint SHA-256: `3c04469072b1e2acf9333ee48ee1002f2db42f46ffb0daab6269686ebc6fe6c1` - Paper: [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - Weights licence: MIT (repo LICENSE; README reserves the paper, diagrams and the name) - Licence note: MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). - Rights review: MIT repository licence; its README reserves the paper, the diagrams and the ST-EEGFormer name, none of which is reproduced. Aggregate scores only. - Row footnote: BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ### ST-EEGFormer Large - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 303.00M - Revision: `official GitHub release large_weights_only_196.pth` - Checkpoint SHA-256: `7b2ed01dec88938ea0a8aba5621eab283bb61ec12b88c6475722fffae8afda83` - Paper: [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - Weights licence: MIT (as Base) - Rights review: MIT repository licence; its README reserves the paper, the diagrams and the ST-EEGFormer name, none of which is reproduced. Aggregate scores only. - Row footnote: BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Notes: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - EEGMAT adaptation: not run by design: ST-EEGFormer Large is frozen probes only ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. ### ST-EEGFormer Base · ST-EEGFormer Large | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://github.com/LiuyinYang1101/STEEGFormer) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://github.com/LiuyinYang1101/STEEGFormer) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://github.com/LiuyinYang1101/STEEGFormer) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://github.com/LiuyinYang1101/STEEGFormer) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://github.com/LiuyinYang1101/STEEGFormer) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | in the authors' published pretraining list | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://openreview.net/forum?id=5Xwm8e6vbh) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | not in the authors' published pretraining list (checked 2026-10-04) | [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) · [source ↗](https://github.com/LiuyinYang1101/STEEGFormer) | **Note on the check:** ST-EEGFormer's in-house data is not described in detail; it is the authors' own lab recordings, so it cannot be any of the seven third-party datasets. ## Cite this page BCI Report (2026). *ST-EEGFormer: results on public EEG datasets*. https://bci.report/methods/st-eegformer/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/st-eegformer/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # xDAWN+LDA and XDAWNCov+TS+LR: results on public EEG datasets Classical method xDAWN is a spatial filter, proposed by Rivet, Souloumiac, Attina and Gibert (IEEE Transactions on Biomedical Engineering, 2009), that enhances an evoked response such as the P300 against the ongoing EEG. BCI Report runs two decoders built on it through pyRiemann with fixed settings (4 filters, OAS covariance estimates): xDAWN+LDA, which flattens the filtered windows and classifies them with a shrinkage linear discriminant, and XDAWNCov+TS+LR, which builds xDAWN-augmented covariance matrices (Barachant and Congedo, 2014), maps them to the tangent space and fits a logistic regression. The settings follow the MOABB reference pipelines (xDAWN_LDA.yml with a shrinkage LDA head; XdawnCov_TS_SVM.yml with the SVM replaced by the logistic regression head). Measured on: [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Protocols: [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/) **Also known as** xDAWN · XDAWNCov · xDAWN+LDA · XDAWNCov+TS+LR · Rivet et al. 2009 **Reference** [Rivet B, Souloumiac A, Attina V, Gibert G. xDAWN algorithm to enhance evoked potentials: application to brain–computer interface. IEEE Transactions on Biomedical Engineering 56(8):2035–2043, 2009. ↗](https://doi.org/10.1109/TBME.2009.2012869) [Official code ↗](https://github.com/pyRiemann/pyRiemann) · Description sources [doi.org](https://doi.org/10.1109/TBME.2009.2012869) · [arxiv.org](https://arxiv.org/abs/1409.0107) · [github.com](https://github.com/pyRiemann/pyRiemann) · [github.com](https://github.com/NeuroTechX/moabb) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds006593](https://bci.report/datasets/ds006593/) · Core matrix · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/) · Chance level 50.0% · [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - XDAWNCov+TS+LR · 4 filters · 50 Hz: 61.9% (58.8%–65.0%) - xDAWN+LDA · 4 filters · 50 Hz: 60.4% (57.7%–63.2%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [XDAWNCov+TS+LR · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Visual target versus nontarget events | Balanced accuracy | 61.9% (58.8%–65.0%) — Primary row · one fixed configuration (MOABB-reference settings) | 21 | | [xDAWN+LDA · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Visual target versus nontarget events | Balanced accuracy | 60.4% (57.7%–63.2%) — Secondary row · one fixed configuration (MOABB-reference settings) | 21 | ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Core matrix · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/) · Chance level 50.0% · [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - XDAWNCov+TS+LR · 4 filters · 50 Hz: 60.3% (58.2%–62.6%) - xDAWN+LDA · 4 filters · 50 Hz: 56.8% (55.1%–58.6%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [XDAWNCov+TS+LR · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Reading · target versus nontarget events | Balanced accuracy | 60.3% (58.2%–62.6%) — Primary row · one fixed configuration (MOABB-reference settings) | 30 | | [xDAWN+LDA · 4 filters · 50 Hz](https://bci.report/methods/xdawn/) | Reading · target versus nontarget events | Balanced accuracy | 56.8% (55.1%–58.6%) — Secondary row · one fixed configuration (MOABB-reference settings) | 30 | [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) Core matrix ## Core-matrix rows: recipe and paired differences Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) ### How these rows were run | Method | Recipe | MOABB reference pipeline | Implementation | Code licence | | --- | --- | --- | --- | --- | | XDAWNCov+TS+LR — Supervised fit | 4 filters · 50 Hz | MOABB XdawnCov_TS_SVM.yml, with the SVM and its grid replaced by the TS+LR head | pyRiemann 0.12 (wheel SHA-256 74a90b2ef0938c94fe9bf90b6506991fc19da55850aae5fdf992026b3845d43a) and scikit-learn | BSD-3-Clause | | xDAWN+LDA — Supervised fit | 4 filters · 50 Hz | MOABB xDAWN_LDA.yml, with shrinkage LDA (lsqr, Ledoit–Wolf) | pyRiemann 0.12 (wheel SHA-256 74a90b2ef0938c94fe9bf90b6506991fc19da55850aae5fdf992026b3845d43a) and scikit-learn | BSD-3-Clause | One fixed configuration per method: library or MOABB-reference settings with the stated deviations. The 4–40 Hz band-pass (or demeaning plus 50 Hz decimation on the ERP protocols) is part of the recipe and is applied to each window on its own. ### Sensitivity lines The same methods with other input filtering, labelled; they are not matrix rows. - **XDAWNCov+TS+LR · 4 filters · raw** · [P300 target ERP](https://bci.report/protocols/p300-target/): 60.7% (57.7%–63.6%), macro F1 0.602, 21 people - **xDAWN+LDA · 4 filters · raw** · [P300 target ERP](https://bci.report/protocols/p300-target/): 58.9% (56.3%–61.5%), macro F1 0.586, 21 people - **XDAWNCov+TS+LR · 4 filters · raw** · [Semantic target ERP](https://bci.report/protocols/semantic-target/): 60.2% (57.8%–62.8%), macro F1 0.589, 30 people - **xDAWN+LDA · 4 filters · raw** · [Semantic target ERP](https://bci.report/protocols/semantic-target/): 54.6% (52.9%–56.4%), macro F1 0.542, 30 people ### Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. Paired participant bootstrap (10,000 draws). The four contrasts against the Spectral ridge cells carry 98.75 % intervals (Bonferroni over four); the other contrasts carry nominal 95 % intervals with no familywise claim. 'Higher' or 'lower' means the interval lies wholly above or below zero; 'unresolved' means it includes zero. ### Primary contrasts against the Spectral ridge cell, with the Temporal ridge beside them | Difference | Mean (pp) | Interval | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | XDAWNCov+TS+LR − Spectral ridge — P300 target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +11.0 pp | 6.8 to 15.1 — 98.75% interval, Bonferroni-adjusted for four contrasts | 7.7 to 14.2 | higher | | XDAWNCov+TS+LR − Temporal ridge — P300 target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +8.1 pp | 5.3 to 11.1 — nominal 95% interval | — | higher | | XDAWNCov+TS+LR − Spectral ridge — Semantic target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +9.3 pp | 6.1 to 12.8 — 98.75% interval, Bonferroni-adjusted for four contrasts | 6.7 to 12.0 | higher | | XDAWNCov+TS+LR − Temporal ridge — Semantic target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +3.9 pp | 2.0 to 5.9 — nominal 95% interval | — | higher | ### Against EEGNet against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | XDAWNCov+TS+LR − EEGNet — P300 target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +8.6 pp | 5.2 to 12.0 | higher | | xDAWN+LDA − EEGNet — P300 target ERP — xDAWN+LDA · 4 filters · 50 Hz | +7.1 pp | 4.3 to 10.0 | higher | | XDAWNCov+TS+LR − EEGNet — Semantic target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | −1.2 pp | −3.0 to 0.6 | unresolved | | xDAWN+LDA − EEGNet — Semantic target ERP — xDAWN+LDA · 4 filters · 50 Hz | −4.6 pp | −6.5 to −2.7 | lower | ### Secondary rows against the Spectral ridge cell | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | xDAWN+LDA − Spectral ridge — P300 target ERP — xDAWN+LDA · 4 filters · 50 Hz | +9.5 pp | 6.5 to 12.4 | higher | | xDAWN+LDA − Spectral ridge — Semantic target ERP — xDAWN+LDA · 4 filters · 50 Hz | +5.8 pp | 4.0 to 7.8 | higher | ### Against the Temporal ridge cell | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | xDAWN+LDA − Temporal ridge — P300 target ERP — xDAWN+LDA · 4 filters · 50 Hz | +6.6 pp | 3.7 to 9.8 | higher | | xDAWN+LDA − Temporal ridge — Semantic target ERP — xDAWN+LDA · 4 filters · 50 Hz | +0.5 pp | −1.2 to 2.1 | unresolved | ### Against the frozen encoders LaBraM and CBraMod | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | XDAWNCov+TS+LR − LaBraM — P300 target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +12.5 pp | 8.7 to 16.3 | higher | | XDAWNCov+TS+LR − CBraMod — P300 target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +6.9 pp | 4.4 to 9.6 | higher | | XDAWNCov+TS+LR − LaBraM — Semantic target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +7.1 pp | 4.8 to 9.4 | higher | | XDAWNCov+TS+LR − CBraMod — Semantic target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +4.5 pp | 2.5 to 6.4 | higher | These pairs use the matrix's LaBraM and CBraMod rows (frozen encoder + ridge head) only. ### Recipe arms: the same method with other input filtering | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | XDAWNCov+TS+LR · 4 filters · 50 Hz − XDAWNCov+TS+LR · 4 filters · raw — P300 target ERP | +1.2 pp | −1.2 to 3.7 | unresolved | | xDAWN+LDA · 4 filters · 50 Hz − xDAWN+LDA · 4 filters · raw — P300 target ERP | +1.5 pp | −0.1 to 2.9 | unresolved | | XDAWNCov+TS+LR · 4 filters · 50 Hz − XDAWNCov+TS+LR · 4 filters · raw — Semantic target ERP | +0.1 pp | −1.3 to 1.3 | unresolved — within two points of zero, sign unresolved | | xDAWN+LDA · 4 filters · 50 Hz − xDAWN+LDA · 4 filters · raw — Semantic target ERP | +2.2 pp | 0.6 to 3.9 | higher | ### Within the family: primary row against secondary row | Difference | Mean (pp) | nominal 95% interval | Reading | | --- | --- | --- | --- | | XDAWNCov+TS+LR − xDAWN+LDA — P300 target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +1.5 pp | −0.9 to 3.7 | unresolved | | XDAWNCov+TS+LR − xDAWN+LDA — Semantic target ERP — XDAWNCov+TS+LR · 4 filters · 50 Hz | +3.5 pp | 1.8 to 5.3 | higher | ### Credits - Rivet, Souloumiac, Attina and Gibert. xDAWN algorithm to enhance evoked potentials: application to brain–computer interface. IEEE Transactions on Biomedical Engineering 56(8), 2009. [doi ↗](https://doi.org/10.1109/TBME.2009.2012869) - Barachant and Congedo. A Plug&Play P300 BCI using information geometry. arXiv:1409.0107, 2014. - Barachant, Bonnet, Congedo and Jutten. Multiclass brain–computer interface classification by Riemannian geometry. IEEE Transactions on Biomedical Engineering 59(4), 2012. [doi ↗](https://doi.org/10.1109/TBME.2011.2172210) - Chen, Wiesel, Eldar and Hero. Shrinkage algorithms for MMSE covariance estimation. IEEE Transactions on Signal Processing 58(10), 2010. [doi ↗](https://doi.org/10.1109/TSP.2010.2053029) - Jayaram and Barachant. MOABB: trustworthy algorithm benchmarking for BCIs. Journal of Neural Engineering 15(6), 2018. [doi ↗](https://doi.org/10.1088/1741-2552/aadea0) - Ledoit and Wolf. A well-conditioned estimator for large-dimensional covariance matrices. Journal of Multivariate Analysis 88(2), 2004. [doi ↗](https://doi.org/10.1016/S0047-259X(03)00096-4) ## Cite this page BCI Report (2026). *xDAWN+LDA and XDAWNCov+TS+LR: results on public EEG datasets*. https://bci.report/methods/xdawn/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/xdawn/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # ZUNA: results on public EEG datasets Foundation model ZUNA 1.1 is a channel-wise EEG encoder from Zyphra. Its authors describe the pretraining data only at repository level — the TUH EEG corpus and OpenNeuro collections — and publish no list of datasets, so whether each core dataset was in it is unknown. BCI Report ran the encoder alone, frozen, with no channel upsampling or reconstruction, and adapted it on EEGMAT. Its model card says research use only, not for diagnosis or clinical use. Measured on: [ds003810](https://bci.report/datasets/ds003810/); [EEGMAT](https://bci.report/datasets/eegmat/); [BETA](https://bci.report/datasets/beta/); [ds006593](https://bci.report/datasets/ds006593/); [TMNRED / ds005383](https://bci.report/datasets/tmnred/); [EESM19 scalp subset](https://bci.report/datasets/eesm19/); [ds005342](https://bci.report/datasets/ds005342/) · Protocols: [Motor imagery & rest](https://bci.report/protocols/mi-rest/); [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/); [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/); [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/); [P300 target ERP](https://bci.report/protocols/p300-target/); [Semantic target ERP](https://bci.report/protocols/semantic-target/); [Sleep staging](https://bci.report/protocols/sleep-scalp/); [Idle & command](https://bci.report/protocols/idle/) **Also known as** ZUNA 1.1 · ZUNA1.1 · Zyphra/ZUNA1.1 **Reference** [ZUNA 1.1, arXiv:2607.27308 ↗](https://arxiv.org/abs/2607.27308) Directory status: Evaluated · [Official code ↗](https://huggingface.co/Zyphra/ZUNA1.1) · Description sources [arxiv.org](https://arxiv.org/abs/2607.27308) · [huggingface.co](https://huggingface.co/Zyphra/ZUNA1.1) ## Published results Compare figures within a group only: task, cohort, chance level and electrode layout can differ between groups. One row is above another only where their 95% intervals do not overlap. ### [ds003810](https://bci.report/datasets/ds003810/) · Further foundation encoders, frozen · motor imagery & rest Source page: [Core-matrix protocol: Motor imagery & rest](https://bci.report/protocols/mi-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 50.7% (49.2%–52.1%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 10 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/mi-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Further foundation encoders, frozen · arithmetic & rest Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) · Chance level 50.0% · [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 53.3% (49.5%–57.1%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [EEGMAT](https://bci.report/datasets/eegmat/) · Model adaptation, further encoders · new people, same task, one fixed recipe Source page: [Core-matrix protocol: Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#v9-adaptation) · Chance level 50.0% · [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) **Balanced accuracy, 2 configurations.** Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level. - ZUNA 1.1 · Frozen encoder + trained head: 51.2% (49.8%–52.7%) - ZUNA 1.1 · LoRA rank 4 + head: 56.9% (53.7%–60.1%) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + trained head | Balanced accuracy | 51.2% (49.8%–52.7%) — 1,218 trainable parameters. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | LoRA rank 4 + head | Balanced accuracy | 56.9% (53.7%–60.1%) — 296,130 trainable parameters. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | LoRA minus frozen + head, same people and folds | Paired difference | +5.7 pp (+2.5 pp–+8.7 pp) — 28 people improved, 7 got worse, 1 unchanged. The interval excludes zero. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 36 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/arithmetic-rest/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 8 channels Source page: [Core-matrix protocol: SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 11.9% (10.8%–13.0%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 70 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/beta-8ch/#v9-licences) ### [BETA](https://bci.report/datasets/beta/) · Further foundation encoders, frozen · SSVEP, 4 channels Source page: [Core-matrix protocol: SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#foundation-v9) · Chance level 2.5% · [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 8.4% (7.5%–9.4%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 70 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/beta-4ch/#v9-licences) ### [ds006593](https://bci.report/datasets/ds006593/) · Further foundation encoders, frozen · P300 target ERP Source page: [Core-matrix protocol: P300 target ERP](https://bci.report/protocols/p300-target/#foundation-v9) · Chance level 50.0% · [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 51.6% (50.3%–53.0%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 21 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/p300-target/#v9-licences) ### [TMNRED / ds005383](https://bci.report/datasets/tmnred/) · Further foundation encoders, frozen · semantic target ERP Source page: [Core-matrix protocol: Semantic target ERP](https://bci.report/protocols/semantic-target/#foundation-v9) · Chance level 50.0% · [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 53.9% (52.2%–55.6%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 30 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/semantic-target/#v9-licences) ### [EESM19 scalp subset](https://bci.report/datasets/eesm19/) · Further foundation encoders, frozen · sleep staging Source page: [Core-matrix protocol: Sleep staging](https://bci.report/protocols/sleep-scalp/#foundation-v9) · Chance level 20.0% · [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + ridge head | Balanced accuracy | 60.1% (57.8%–62.3%) — Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 20 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/sleep-scalp/#v9-licences) ### [ds005342](https://bci.report/datasets/ds005342/) · Further foundation encoders, frozen · idle & command Source page: [Core-matrix protocol: Idle & command](https://bci.report/protocols/idle/#foundation-v9) · [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | Method | Condition | Metric | Value | People | | --- | --- | --- | --- | --- | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + linear head | Commands detected ≤3 s | 11 / 60 — 0 of 4 people always abstained. Pretraining exposure: unknown: the authors do not list their pretraining data closely enough to decide. Model card: research use only, not for diagnosis or clinical use. | 4 | | [ZUNA 1.1](https://bci.report/methods/zuna/) | Frozen encoder + linear head | Idle false activations | 1 / 60 — Model card: research use only, not for diagnosis or clinical use. | 4 | Weights terms: ZUNA under Apache-2.0. [Every weights licence and its terms →](https://bci.report/protocols/idle/#v9-licences) [All methods →](https://bci.report/methods/) · [All datasets →](https://bci.report/datasets/) ## Checkpoints, inputs and terms No model’s authors endorse these results. No weights, adapted weights or LoRA deltas were kept or shared. ### ZUNA 1.1 Model card: research use only, not for diagnosis or clinical use. - Panel: A row beside the core matrix on every protocol page - Encoder parameters: 172.10M - Revision: `Zyphra/ZUNA1.1 @ 10b9d6c8` - Checkpoint SHA-256: `295fcddb42ab4623dc1346fba12176798d97ab95389c666907e32cbce23f7f33` - Paper: [paper ↗](https://arxiv.org/abs/2607.27308) - Weights licence: Apache-2.0; model card: research use only, not for diagnosis or clinical use - Licence note: Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. - Rights review: Apache-2.0 weights. The model card says research use only, not for diagnosis or clinical use; that sentence travels with its rows. - Row footnote: Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - Notes: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - EEGMAT adaptation: Run: a head trained on the frozen encoder against LoRA rank 4, three seeds. ## Pretraining exposure, by dataset Whether each core dataset is in the model authors’ published pretraining list, checked on 2026-10-04 at dataset level and linked to the source; a dataset absent from the list may still overlap the pretraining data. | Dataset | Protocols | Statement | Source | | --- | --- | --- | --- | | [ds003810 Motor Imagery vs Rest (low-cost EEG; Peterson et al.)](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://github.com/Zyphra/zuna/blob/main/src/zuna/inference/AY2l/lingua/apps/AY2latent_bci/configs/config_bci_eval.yaml) · [source ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) | | [EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)](https://physionet.org/content/eegmat/1.0.0/) | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://openneuro.org/crn/graphql) · [source ↗](https://physionet.org/content/eegmat/1.0.0/) | | [ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | [P300 target ERP](https://bci.report/protocols/p300-target/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://github.com/Zyphra/zuna/blob/main/src/zuna/inference/AY2l/lingua/apps/AY2latent_bci/configs/config_bci_eval.yaml) · [source ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) | | [ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://github.com/Zyphra/zuna/blob/main/src/zuna/inference/AY2l/lingua/apps/AY2latent_bci/configs/config_bci_eval.yaml) · [source ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) | | [EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://github.com/Zyphra/zuna/blob/main/src/zuna/inference/AY2l/lingua/apps/AY2latent_bci/configs/config_bci_eval.yaml) · [source ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) | | [BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)](https://figshare.com/articles/dataset/The_BETA_database/12264401) | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) · [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://openneuro.org/crn/graphql) · [source ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) | | [ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | [Idle & command](https://bci.report/protocols/idle/) | unknown: the authors do not list their pretraining data closely enough to decide | [source ↗](https://arxiv.org/abs/2602.18478) · [source ↗](https://arxiv.org/abs/2607.27308) · [source ↗](https://github.com/Zyphra/zuna/blob/main/src/zuna/inference/AY2l/lingua/apps/AY2latent_bci/configs/config_bci_eval.yaml) · [source ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) | ## Cite this page BCI Report (2026). *ZUNA: results on public EEG datasets*. https://bci.report/methods/zuna/ Figures from release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04). Cite the upstream datasets as well: each dataset’s page gives its credit. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/methods/zuna/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # The protocols behind the core matrix Protocols Each protocol fixes who was tested, how the data were split, which electrodes and window a method saw, what it was allowed to learn, and what guessing would score. ## Core-matrix protocols | Protocol | Dataset | Task | Evaluation | People | Chance level | Methods scored | | --- | --- | --- | --- | --- | --- | --- | | [Motor imagery & rest](https://bci.report/protocols/mi-rest/) | [ds003810](https://bci.report/datasets/ds003810/) | Rest versus right-hand imagery | Transfer to a new person | 10 | 50.0% | Spectral ridge, [CSP+LDA](https://bci.report/methods/csp-lda/), [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/) — [+ 13 further foundation encoders, frozen →](https://bci.report/protocols/mi-rest/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/mi-rest/#eegpt) | | [Idle & command](https://bci.report/protocols/idle/) | [ds005342](https://bci.report/datasets/ds005342/) | Seated motor imagery · cue-gated replay | False activation × detection | 4 | none — detection is read with false activations | [CSP+LDA](https://bci.report/methods/csp-lda/), [LaBraM](https://bci.report/methods/labram/), [EEGNet](https://bci.report/methods/eegnet/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [CBraMod](https://bci.report/methods/cbramod/), [Deep4Net](https://bci.report/methods/deep4net/), [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/) — [+ 13 further foundation encoders, frozen →](https://bci.report/protocols/idle/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/idle/#eegpt) | | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/) | [BETA](https://bci.report/datasets/beta/) | 40 visual targets · 8 posterior electrodes | Transfer to a new person | 70 | 2.5% | [Standard CCA](https://bci.report/methods/cca/), Spectral ridge, [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [FBCCA](https://bci.report/methods/fbcca/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/) — [+ 13 further foundation encoders, frozen →](https://bci.report/protocols/beta-8ch/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/beta-8ch/#eegpt) | | [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/) | [BETA](https://bci.report/datasets/beta/) | 40 visual targets · 4 posterior electrodes | Transfer to a new person | 70 | 2.5% | [Standard CCA](https://bci.report/methods/cca/), Spectral ridge, [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [FBCCA](https://bci.report/methods/fbcca/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/) — [+ 13 further foundation encoders, frozen →](https://bci.report/protocols/beta-4ch/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/beta-4ch/#eegpt) | | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/) | [EEGMAT](https://bci.report/datasets/eegmat/) | Serial subtraction versus resting EEG | Transfer to a new person | 36 | 50.0% | Spectral ridge, [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/), [CSP+LDA](https://bci.report/methods/csp-lda/) — [+ 13 further foundation encoders, frozen →](https://bci.report/protocols/arithmetic-rest/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/arithmetic-rest/#eegpt) | | [P300 target ERP](https://bci.report/protocols/p300-target/) | [ds006593](https://bci.report/datasets/ds006593/) | Visual target versus nontarget events | Transfer to a new person | 21 | 50.0% | Spectral ridge, Temporal ridge, [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/), [CSP+LDA](https://bci.report/methods/csp-lda/) — [+ 12 further foundation encoders, frozen →](https://bci.report/protocols/p300-target/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/p300-target/#eegpt) | | [Semantic target ERP](https://bci.report/protocols/semantic-target/) | [TMNRED / ds005383](https://bci.report/datasets/tmnred/) | Reading · target versus nontarget events | Transfer to a new person | 30 | 50.0% | Spectral ridge, Temporal ridge, [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/), [CSP+LDA](https://bci.report/methods/csp-lda/) — [+ 12 further foundation encoders, frozen →](https://bci.report/protocols/semantic-target/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/semantic-target/#eegpt) | | [Sleep staging](https://bci.report/protocols/sleep-scalp/) | [EESM19 scalp subset](https://bci.report/datasets/eesm19/) | Five stages · balanced scalp-EEG sample | Transfer to a new person | 20 | 20.0% | Spectral ridge, [LaBraM](https://bci.report/methods/labram/), [CBraMod](https://bci.report/methods/cbramod/), [EEGNet](https://bci.report/methods/eegnet/), [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/), [Deep4Net](https://bci.report/methods/deep4net/), [CSP+LDA](https://bci.report/methods/csp-lda/) — [+ 13 further foundation encoders, frozen →](https://bci.report/protocols/sleep-scalp/#foundation-v9) — [+ EEGPT, frozen →](https://bci.report/protocols/sleep-scalp/#eegpt) | In the matrix, a dash marks a method that does not apply to the protocol, and a dot a method not run on it. [The core matrix on the home page →](https://bci.report/#overview) · [Core matrix · JSON ↓](https://bci.report/data/experiments.json) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) --- Markdown copy of https://bci.report/protocols/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Motor imagery & rest on ds003810 Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: Rest versus right-hand imagery - Evaluation: Transfer to a new person - Dataset: [ds003810](https://bci.report/datasets/ds003810/) - People: 10 - Data: 1,200 epochs · 5 participant-disjoint folds - Input: 15 channels · 2-second windows - Chance level: 50.0% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Spectral ridge: 54.9% (53.2%–56.7%) - CSP+LDA: 47.1% (43.8%–50.2%) - LaBraM: 53.5% (51.4%–55.4%) - CBraMod: 62.9% (59.6%–66.5%) - EEGNet: 71.0% (65.0%–76.3%) - TS+LR: 58.9% (54.9%–63.9%) - MDM: 53.8% (50.8%–57.9%) - ShallowFBCSPNet: 68.8% (63.4%–73.9%) - Deep4Net: 68.6% (63.3%–74.0%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method | Supervised fit | 54.9% (53.2%–56.7%) | 0.544 — Mean across held-out participants | 15 | 10 | 0.2 s | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method | Supervised fit | 47.1% (43.8%–50.2%) At or below chance level | 0.406 — Mean across held-out participants | 15 | 10 | 8.8 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 53.5% (51.4%–55.4%) | 0.509 — Mean across held-out participants | 15 | 10 | 2.2 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 62.9% (59.6%–66.5%) | 0.617 — Mean across held-out participants | 15 | 10 | 5.0 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 71.0% (65.0%–76.3%) Single seed — the highest of the seeds run (see Stability) | 0.680 — Mean across held-out participants | 15 | 10 | 55.5 s | | [TS+LR](https://bci.report/methods/riemannian/) — Classical method · OAS · 4–40 Hz · primary row · Local Ubuntu / CPU | Supervised fit | 58.9% (54.9%–63.9%) | 0.534 — Mean across held-out participants | 15 | 10 | — | | [MDM](https://bci.report/methods/riemannian/) — Classical method · OAS · 4–40 Hz · secondary row · Local Ubuntu / CPU | Supervised fit | 53.8% (50.8%–57.9%) | 0.420 — Mean across held-out participants | 15 | 10 | — | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 68.8% (63.4%–73.9%) | 0.672 — Mean across held-out participants | 15 | 10 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 68.6% (63.3%–74.0%) | 0.659 — Mean across held-out participants | 15 | 10 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [Standard CCA](https://bci.report/methods/cca/), [FBCCA](https://bci.report/methods/fbcca/) (they need the stimulation frequencies, and this protocol has none); Temporal ridge, [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/) (they model the waveform of a stimulus-locked transient potential, and this protocol's classes are not defined by one). [A frozen NeuroRVQ probe with its own input chain is on the pilot studies page →](https://bci.report/topics/pilot-studies/#neurorvq) ## How these rows were run ### Riemannian and xDAWN baselines - Recipe: TS+LR: OAS · 4–40 Hz · primary row; MDM: OAS · 4–40 Hz · secondary row - Hardware: TS+LR, MDM: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) One fixed configuration per method: library or MOABB-reference settings with the stated deviations. The 4–40 Hz band-pass (or demeaning plus 50 Hz decimation on the ERP protocols) is part of the recipe and is applied to each window on its own. - **TS+LR**: Primary row · one fixed configuration (MOABB-reference settings) - **MDM**: Secondary row · one fixed configuration (MOABB-reference settings) ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW - Hardware: ShallowFBCSPNet, Deep4Net: Apple M5 / MPS - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### Riemannian and xDAWN baselines | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | TS+LR − Spectral ridge — TS+LR · OAS · 4–40 Hz | +4.0 pp | −1.1 to 11.5 | −0.3 to 9.7 | unresolved | Paired participant bootstrap (10,000 draws). The four contrasts against the Spectral ridge cells carry 98.75 % intervals (Bonferroni over four); the other contrasts carry nominal 95 % intervals with no familywise claim. 'Higher' or 'lower' means the interval lies wholly above or below zero; 'unresolved' means it includes zero. Ten people: a percentile bootstrap over ten people tends to give intervals narrower than their nominal level. ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | Deep4Net − EEGNet | −2.4 pp | −6.8 to 2.3 | −8.6 to 5.2 | unresolved | | ShallowFBCSPNet − EEGNet | −2.2 pp | −8.0 to 3.4 | −10.8 to 5.7 | unresolved | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 54.92% (53.17%–56.67%) | 0.544 | 15 | 10 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 57.33% (54.42%–60.33%) | 0.558 | 15 | 10 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: +2.42 pp (−1.17 to +5.83): not separated. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus SingLEM, the highest of the 14 frozen-encoder rows this comparison was specified against, chosen on these same test folds: −14.67 pp (−19.33 to −9.75): below. Choosing it on the test folds leans the comparison against this row. ### Notes on this comparator The band-power rows in the pilot studies panels on this cohort come from their own input chains and are read on their own. [On these windows and folds this is the same comparator as the third block of the panel on the components of Neural State Prediction on the pilot studies page →](https://bci.report/topics/pilot-studies/#nsp-components-a) - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 54.9% (53.2%–56.7%) - CSP+LDA: 47.1% (43.8%–50.2%) - LaBraM: 53.5% (51.4%–55.4%) - CBraMod: 62.9% (59.6%–66.5%) - EEGNet: 71.0% (65.0%–76.3%) - REVE Base: 66.2% (61.8%–70.6%) - REVE Large: 70.2% (64.0%–76.1%) - LUNA Base: 52.6% (50.2%–54.9%) - LUNA Large: 49.9% (46.6%–53.5%) - BrainOmni Base: 58.8% (57.0%–60.8%) - CodeBrain: 58.8% (54.5%–62.6%) - EEGMamba: 67.3% (63.6%–70.6%) - ST-EEGFormer Base: 58.1% (53.7%–63.3%) - ST-EEGFormer Large: 62.4% (57.4%–67.8%) - eeg-fm-masking MAE (r=9 cm, L=2): 68.1% (64.3%–72.3%) - ERP-FM Base: 54.8% (51.8%–57.8%) - SingLEM: 72.0% (66.6%–76.6%) - ZUNA 1.1: 50.7% (49.2%–52.1%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 66.2% (61.8%–70.6%) | 0.640 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 70.2% (64.0%–76.1%) | 0.673 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 52.6% (50.2%–54.9%) | 0.503 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 49.9% (46.6%–53.5%) At or below chance level | 0.486 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | 58.8% (57.0%–60.8%) | 0.582 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 58.8% (54.5%–62.6%) | 0.573 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 67.3% (63.6%–70.6%) | 0.669 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 58.1% (53.7%–63.3%) | 0.541 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 62.4% (57.4%–67.8%) | 0.586 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 68.1% (64.3%–72.3%) | 0.677 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 54.8% (51.8%–57.8%) | 0.524 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 72.0% (66.6%–76.6%) | 0.715 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 50.7% (49.2%–52.1%) Interval reaches chance level | 0.499 | 15 | 10 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2602.18478) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 68.1% (64.3%–72.3%) | 0.677 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 63.2% (58.2%–67.9%) | 0.620 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 64.9% (60.3%–69.4%) | 0.645 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 63.3% (59.0%–67.4%) | 0.628 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. Point estimate at chance; reported as is. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - **ERP-FM Base**: Outside the ERP-FM design: negative control. - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-mi-rest.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 54.9% (53.2%–56.7%) - CSP+LDA: 47.1% (43.8%–50.2%) - LaBraM: 53.5% (51.4%–55.4%) - CBraMod: 62.9% (59.6%–66.5%) - EEGNet: 71.0% (65.0%–76.3%) - EEGPT: 58.3% (54.9%–61.8%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 58.3% (54.9%–61.8%) | 0.555 | 15 | 10 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - ds003810 is recorded at 125 Hz; its windows reach the encoder upsampled to 256 Hz, so they carry nothing above 62.5 Hz. - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Spectral ridge: the intervals overlap - CSP+LDA: EEGPT above, the intervals 4.67 pp apart - LaBraM: the intervals overlap - CBraMod: the intervals overlap - EEGNet: EEGPT below, the intervals 3.25 pp apart EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 61.9% (59.1%–64.4%) | 0.612 | — | | Quark-L — Local Ubuntu / CPU | 59.6% (56.3%–62.9%) | 0.578 | — | | Quark-H — Local Ubuntu / CPU | 60.2% (57.4%–62.9%) | 0.596 | — | | Quark-G † — Local Ubuntu / CUDA | 58.7% (57.3%–60.1%) | 0.579 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 56.9% (53.3%–61.0%) | 0.539 | its pretraining corpus includes this protocol's dataset | | Molecule-L — Local Ubuntu / CPU | 59.6% (57.2%–61.4%) | 0.577 | its pretraining corpus includes this protocol's dataset | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 61.5% (58.3%–64.8%) | 0.590 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) Two more models · this protocol’s people and folds ## Signal-JEPA and EEGSimpleConv Signal-JEPA as a frozen encoder with a ridge head fitted on the training people, and EEGSimpleConv trained from scratch for 20 epochs, both on this protocol’s windows, people and 5 folds. Each is compared with EEGNet on the same people. | Model | Training mode | Input | Balanced accuracy | Macro F1 | Difference from EEGNet | | --- | --- | --- | --- | --- | --- | | Signal-JEPA — Foundation model · 3,456,852 parameters | Frozen encoder + ridge head | 0.5–40 Hz band-pass, 128 Hz | 67.25% (63.42%–70.50%) | 0.669 | −3.75 pp (−7.58 to −0.17) — vs this page’s EEGNet cell: 71.0% — against the matrix's one-seed EEGNet cell; training variance not included | | EEGSimpleConv — Compact model · 185,022 parameters | Trained from scratch, 20 epochs | the matrix’s 250 Hz windows | 59.33% — mean of three seeds | 0.530 | −10.14 pp — vs EEGNet’s three-seed mean · no interval | Each difference is a paired comparison with EEGNet on the same people. ### Notes on Signal-JEPA and EEGSimpleConv Signal-JEPA’s input: each window band-passed 0.5–40 Hz and resampled to 128 Hz, the format it was pretrained on, in microvolts, recorded channels only. The further-encoder panel follows its own input contract, so this row sits apart from it. Signal-JEPA’s difference is paired, on the same people, against this page’s EEGNet cell: 10,000 participant bootstrap draws, no multiplicity correction. EEGNet’s mean over its three seeds here is 69.47%. EEGSimpleConv: the mean of three seeds, at width 64 and the matrix’s 20-epoch budget, fixed before scoring. Its difference is from EEGNet’s mean over the same three seeds and is printed without an interval. Signal-JEPA’s model card names Lee2019 (OpenBMI) as its pretraining data. ### Terms of Signal-JEPA and EEGSimpleConv - Signal-JEPA: weights MIT, as the model card at revision 51232ee0 states; code BSD-3-Clause (braindecode 1.5.1); the weights are not redistributed. [paper ↗](https://arxiv.org/abs/2403.11772) · [model card ↗](https://huggingface.co/braindecode/signal-jepa) - EEGSimpleConv: code BSD-3-Clause (braindecode 1.5.1); the separate author repository’s licence is unclear; trained from scratch, and no trained weights are shared. [paper ↗](https://arxiv.org/abs/2309.07159) - [Signal-JEPA and EEGSimpleConv · JSON ↓](https://bci.report/data/signal-jepa-eegsimpleconv-update.json) - [Release `signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/#signal-jepa-eegsimpleconv-update-20261106) ## Limits Ten-person laboratory task with prompted rest. This is not continuous-idle monitoring or a physical low-channel headset test. ### The protocol, step by step - 5 participant-disjoint folds. All recordings from a person stay together. Each person contributes to the held-out predictions once. - epoching: two seconds from the class annotation onset; amplitude: converted to microvolts according to source calibration, then per-channel epoch mean removed; resampling: none; native sampling rate retained; selection: natural file/event order; when over the cap, retain 60 evenly spaced event indices per participant/class; source_units: Microv declared by BIDS channels.tsv; EDF physical dimension is absent, so MNE returns the source numeric microvolt values without SI scaling - One fixed seed (20260919); no early stopping or test-based tuning. EEGNet trains for 20 epochs per fold. Frozen encoders use training-only standardized ridge heads (alpha 100). - Labels: rest, right_hand_imagery. - Uniform-guessing reference: 50.00%. Scores weight participants equally. Intervals describe participant variation; cross-validation training sets overlap. - run 0 is real dominant-hand movement and is excluded; only imagery runs 1-4 contribute - the source reports online 0.5-45 Hz filtering - controlled cue-locked laboratory windows; this does not measure continuous false activations - foundation-model pretraining overlap is unknown - EEGNet three-seed mean 69.47%; sample SD 1.43 percentage points; range 68.17–71.00%. Main table retains the original fixed seed; this is not a confidence interval. ### Stability EEGNet three-seed mean 69.47%; sample SD 1.43 percentage points; range 68.17–71.00%. Main table retains the original fixed seed; this is not a confidence interval. Seeds run for EEGNet: 71.00% · 69.25% · 68.17%; mean 69.47% Same participants, folds, preprocessing and 20-epoch budget. Three seeds measure initialization variability, not population uncertainty. Main table retains its preselected seed; no best-seed selection. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. EEGNet three-seed mean 69.47%; sample SD 1.43 percentage points; range 68.17–71.00%. Main table retains the original fixed seed; this is not a confidence interval. ### Model terms - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **CSP+LDA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **TS+LR**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **MDM**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### ds003810 **Credit** Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ **The dataset authors ask you to cite:** Peterson V, Galván C, Hernández H, Spies R. A feasibility study of a complete low-cost consumer-grade brain-computer interface system. Heliyon 6(3):e03425 (2020). [doi ↗](https://doi.org/10.1016/j.heliyon.2020.e03425) · [as their record asks ↗](https://openneuro.org/datasets/ds003810) **Licence** [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) [Dataset record ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) · [Every result on this dataset →](https://bci.report/datasets/ds003810/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/mi-rest-protocol.json) **Public-data register note** Clear CC0 snapshot plus dataset-specific ethics and signed-consent evidence; aggregate-only publication materially limits privacy exposure. **Rights reviewed** 2026-09-20 · against [doi.org ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) · [pmc.ncbi.nlm.nih.gov ↗](https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/) **Correction on record · 2026-10-01**: The ds003810 credit links only the 2022 Data in Brief description. The OpenNeuro record asks users to cite Peterson, Galván, Hernández and Spies, Heliyon 6(3):e03425 (2020); the dataset page now gives both. [Corrections register →](https://bci.report/releases/#corrections) ## Reproducibility record - Protocol id: `parallel-fixed-subject-folds-v1/ds003810` - Dataset release: OpenNeuro snapshot 2.0.2 · c674303319303b9ffc4ff7e2340b3a4807e8ffb5 - Hardware: Apple M5 / MPS - Scoring stage sum: 71.7 s · may include accelerator waiting - Audit record SHA-256: `e762cf7f048bf2df47b9399b9df6171733a93c162ce7800560952df30491a7f2` - Summary SHA-256: `0dd66f4bd4d083fca1d1ab44a2d2951d57f6d19378079b8cb5ff3ee9241f0cda` - Protocol SHA-256: `57ddb7692e3bb248b95c4e81b6aa02e7c8b19d774ed4a8b799c811b6ef3b65a2` ### Rows of core-matrix-update-20261106 #### Riemannian and xDAWN baselines - Protocol SHA-256: `4310329a2d6dddaeff64a965750b64237b08220fc23717b1f3d0c1a48f56b055` - Audit record SHA-256: `ce1ee2a0345358c613aa89298c1cb90939df15d68d6c48298b99f9daa7f69e82` - Hardware: Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws #### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Apple M5 / MPS - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/mi-rest-results.csv) - [Protocol · JSON ↓](https://bci.report/data/mi-rest-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-mi-rest.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Motor imagery & rest on ds003810*. https://bci.report/protocols/mi-rest/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/#signal-jepa-eegsimpleconv-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/mi-rest/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Idle & command on ds005342 Core-matrix protocol · False activation × detection The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: Seated motor imagery · cue-gated replay - Evaluation: False activation × detection - Dataset: [ds005342](https://bci.report/datasets/ds005342/) - People: 4 - Data: 60 idle / 60 command test trials - Input: 180 seconds of test idle - Chance level: Not applicable. Command detection is read together with the idle false-activation rate: always abstaining scores zero false activations. - Metrics: Command detection ≤3s (primary) · Idle false activation (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. | Method | Training mode | Command detection ≤3s | Idle false activation | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method | Supervised fit | 23.3% — 14 / 60 command trials | 0.0% — 0 / 60 idle trials 1 of 4 people always abstained | 17 | 4 | 0.1 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + linear head | 18.3% — 11 / 60 command trials | 3.3% — 2 / 60 idle trials — 0 of 4 people always abstained | 17 | 4 | 2.8 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 1.7% — 1 / 60 command trials | 0.0% — 0 / 60 idle trials 2 of 4 people always abstained | 17 | 4 | 10.1 s | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model | Scratch · 20 epochs | 43.3% — 26 / 60 command trials | 1.7% — 1 / 60 idle trials 1 of 4 people always abstained | 17 | 4 | 28.6 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + linear head | 18.3% — 11 / 60 command trials | 1.7% — 1 / 60 idle trials 1 of 4 people always abstained | 17 | 4 | 2.3 s | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model | Scratch · 20 epochs | 3.3% — 2 / 60 command trials | 1.7% — 1 / 60 idle trials 1 of 4 people always abstained | 17 | 4 | 13.9 s | | [TS+LR](https://bci.report/methods/riemannian/) — Classical method · OAS · idle-native windows · Apple M5 / CPU | Supervised fit | 35.0% — 21 / 60 command trials | 6.7% — 4 / 60 idle trials | 17 | 4 | — | | [MDM](https://bci.report/methods/riemannian/) — Classical method · OAS · idle-native windows · Apple M5 / CPU | Supervised fit | 35.0% — 21 / 60 command trials | 1.7% — 1 / 60 idle trials | 17 | 4 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [Standard CCA](https://bci.report/methods/cca/), [FBCCA](https://bci.report/methods/fbcca/) (they need the stimulation frequencies, and this protocol has none); Temporal ridge, [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/) (they model the waveform of a stimulus-locked transient potential, and this protocol's classes are not defined by one). **Not run under this protocol:** Spectral ridge. ## How these rows were run ### Riemannian and xDAWN baselines - Recipe: TS+LR, MDM: OAS · idle-native windows - Protocol: For each subject, the first two blocks train the model, the third selects the threshold, and the final block is held out for testing. - Selection: One fixed configuration per method: library or MOABB-reference settings with the stated deviations. - Hardware: TS+LR, MDM: Apple M5 / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) - **TS+LR, MDM**: Summed test trials. Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. - **MDM**: MDM's probabilities are a softmax of minus the squared Riemannian distance, an uncalibrated scale, read against the published threshold grid (0.50–0.99). - **People who always abstained** (TS+LR, MDM): Left empty: counts of people meeting a criterion print only for cohorts of 10 or more. #### Idle: differences in summed test trials Differences in summed test-trial counts against the published idle rows; descriptive, no interval. Against EEGNet: against the matrix's one-seed EEGNet cell; training variance not included. Against CSP+LDA: against the published CSP+LDA idle row (deterministic). | Method | Against EEGNet · Idle trials falsely triggered | Against EEGNet · Command trials detected | Against CSP+LDA · Idle trials falsely triggered | Against CSP+LDA · Command trials detected | | --- | --- | --- | --- | --- | | TS+LR | +4 | +20 | +4 | +7 | | MDM | +1 | +20 | +1 | +7 | Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, per-person linear head with its calibration-only threshold and scoring stay the same. Rows are grouped by model family; they are counts with no interval, so no row is called above or below another. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. | Encoder | Training mode | Command detection ≤3 s | Idle false activation | Window balanced accuracy · AUROC | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + linear head | 20.0% — 12 of 60 commands — 0 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 61.1% — AUROC 0.786 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + linear head | 10.0% — 6 of 60 commands 2 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 54.4% — AUROC 0.785 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + linear head | 6.7% — 4 of 60 commands 2 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 52.8% — AUROC 0.675 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + linear head | 8.3% — 5 of 60 commands 2 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 54.7% — AUROC 0.734 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + linear head | 21.7% — 13 of 60 commands 1 of 4 people always abstained | 5.0% — 3 of 60 idle trials | 59.7% — AUROC 0.737 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + linear head | 16.7% — 10 of 60 commands 1 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 58.6% — AUROC 0.771 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + linear head | 3.3% — 2 of 60 commands 3 of 4 people always abstained | 0.0% — 0 of 60 idle trials | 51.1% — AUROC 0.639 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + linear head | 26.7% — 16 of 60 commands 1 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 63.9% — AUROC 0.896 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + linear head | 33.3% — 20 of 60 commands 1 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 67.8% — AUROC 0.912 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + linear head | 5.0% — 3 of 60 commands 2 of 4 people always abstained | 0.0% — 0 of 60 idle trials | 53.6% — AUROC 0.733 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + linear head | 5.0% — 3 of 60 commands 2 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 52.5% — AUROC 0.675 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + linear head | 10.0% — 6 of 60 commands 2 of 4 people always abstained | 8.3% — 5 of 60 idle trials | 50.3% — AUROC 0.528 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + linear head | 18.3% — 11 of 60 commands — 0 of 4 people always abstained | 1.7% — 1 of 60 idle trials | 57.2% — AUROC 0.608 | 17 | 4 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2602.18478) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Command detection ≤3 s | Idle false activation | Window balanced accuracy · AUROC | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + linear head | 5.0% — 3 of 60 commands 2 of 4 people always abstained | 0.0% — 0 of 60 idle trials | 53.6% — AUROC 0.733 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + linear head | 20.0% — 12 of 60 commands 2 of 4 people always abstained | 0.0% — 0 of 60 idle trials | 62.5% — AUROC 0.823 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + linear head | 8.3% — 5 of 60 commands 2 of 4 people always abstained | 0.0% — 0 of 60 idle trials | 57.2% — AUROC 0.768 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + linear head | 8.3% — 5 of 60 commands 2 of 4 people always abstained | 0.0% — 0 of 60 idle trials | 54.7% — AUROC 0.795 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - **ERP-FM Base**: Outside the ERP-FM design: negative control. - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-idle.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same run as the published LaBraM row above, with only the encoder swapped: the same windows, people, folds, per-person logistic head with its calibration-only rule and scoring. Its figures are counts and cohort means with no interval, so it is not called above or below another row. | Encoder | Training mode | Command detection ≤3 s | Idle false activation | Window balanced accuracy · AUROC | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + linear head | 11.7% — 7 of 60 commands 1 of 4 people always abstained | 3.3% — 2 of 60 idle trials | 57.2% — AUROC 0.780 | 17 | 4 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - Same run as the published LaBraM idle row; the published CBraMod idle row comes from another run, so a comparison with it is across runs. The window balanced accuracy and AUROC were computed from this run’s saved probabilities with the definitions of the further foundation-encoder rows; the run itself recorded the counts. - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Cohort size n = 4 (a protocol fact): idle trials with a false trigger, command trials detected within 3 s, and the window metrics. | Encoder | Idle false triggers | Commands detected ≤3 s | Mean window balanced accuracy | Mean window AUROC | | --- | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | | Quark-B — Local Ubuntu / CPU | 4 / 60 | 15 / 60 | 60.8% | 0.724 | | Quark-L — Local Ubuntu / CPU | 2 / 60 | 13 / 60 | 61.1% | 0.655 | | Quark-H — Local Ubuntu / CPU | 2 / 60 | 15 / 60 | 62.8% | 0.746 | | Quark-G — Local Ubuntu / CUDA | 2 / 60 | 10 / 60 | 57.2% | 0.722 | | LaBraM at Scale · Atom-L and Molecule-L | | | | | | Atom-L — Local Ubuntu / CPU | 2 / 60 — its pretraining corpus includes this protocol's dataset | 20 / 60 | 66.1% | 0.721 | | Molecule-L — Local Ubuntu / CPU | 1 / 60 — its pretraining corpus includes this protocol's dataset | 10 / 60 | 60.6% | 0.723 | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | | REVE Small — Local Ubuntu / CPU | 1 / 60 | 8 / 60 | 54.4% | 0.680 | Idle is a within-person protocol: each row gives the cohort size, trial counts and window metrics, with no contrast. LaBraM at Scale's pooled features are global on idle; the published LaBraM idle row kept the channel axis. Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Limits Detection resets at each cue. Only three minutes of test idle are observed; these results cannot estimate natural, always-on false activations per hour. ### The protocol, step by step - For each subject, the first two blocks train the model, the third selects the threshold, and the final block is held out for testing. - Class events 1/2 define task onset. Two-second history windows produce decisions at 2.0, 2.5 and 3.0 seconds. - Two consecutive threshold crossings trigger one command. Thresholds use calibration data only and may reject every command. - 17 channels at 250 Hz; one-way 4–40 Hz filtering. Normalization uses training data only; no future samples enter a decision window. - Visual cues, original feedback and short calibration exposure can influence results. This is not a natural continuous-idle test. - Additional models use protocol ds005342-sitting-cue-gated-expanded-mvp-v1; independent audit SHA256: 14445d3423b5268654a1f73f7e1aca8bb03a2f793b076f450883fea89c7bcc83. ### Stability Single fixed configuration; no test-set hyperparameter selection. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - Fixed calibration rule; 60 training trials and 30 calibration trials per participant. Four-person case study; cohort aggregates only, no subgroup or population claim. ### Model terms - **CSP+LDA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **TS+LR**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **MDM**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. ## Source and licence ### ds005342 **Credit** OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. **Licence** [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) · [Every result on this dataset →](https://bci.report/datasets/ds005342/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/idle-protocol.json) **Public-data register note** Eligible only as an aggregate-only case study: remove all participant IDs/results and condition breakdowns, label n=4 prominently, avoid subgroup claims, and make no claim that the cohort is anonymous. **Rights reviewed** 2026-09-20 · against [doi.org ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) · [eegdash.org ↗](https://eegdash.org/api/dataset/eegdash.dataset.DS005342.html) · [doi.org ↗](https://doi.org/10.3389/fninf.2022.961089) ## Reproducibility record - Protocol id: `ds005342-sitting-cue-gated-pilot-v1` - Dataset release: OpenNeuro snapshot 1.0.3 · 30bb3b67d76ce007b822fed801e496fafe5598a0 - Hardware: Apple M5 / MPS and CPU - Scoring stage sum: 58.0 s · may include accelerator waiting - Audit record SHA-256: `ce2474c547f5f11c256752b0a07466beb5b6928da34bb39783aaa4647973d2e4` ### Rows of core-matrix-update-20261106 #### Riemannian and xDAWN baselines - Protocol SHA-256: `4310329a2d6dddaeff64a965750b64237b08220fc23717b1f3d0c1a48f56b055` - Audit record SHA-256: `ce1ee2a0345358c613aa89298c1cb90939df15d68d6c48298b99f9daa7f69e82` - Hardware: Apple M5 / CPU ## Downloads - [Results · CSV ↓](https://bci.report/data/idle-results.csv) - [Protocol · JSON ↓](https://bci.report/data/idle-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-idle.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Idle & command on ds005342*. https://bci.report/protocols/idle/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/idle/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # SSVEP · 8 channels on BETA Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: 40 visual targets · 8 posterior electrodes - Evaluation: Transfer to a new person - Dataset: [BETA](https://bci.report/datasets/beta/) - People: 70 - Data: 11,200 trials · 7 participant-disjoint folds - Input: 8 channels selected from a 64-channel recording · 2-second windows - Chance level: 2.5% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Standard CCA: 63.1% (57.2%–69.0%) - Spectral ridge: 50.7% (45.6%–55.8%) - LaBraM: 10.8% (9.5%–12.2%) - CBraMod: 33.7% (29.8%–37.6%) - EEGNet: 55.8% (50.1%–61.3%) - FBCCA: 79.1% (75.1%–82.8%) - ShallowFBCSPNet: 23.0% (20.3%–25.9%) - Deep4Net: 52.8% (47.3%–58.2%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) — Classical method | Known-frequency reference · no fitting | 63.1% (57.2%–69.0%) | 0.626 — Mean across held-out participants | 8 | 70 | 1.8 s | | Spectral ridge — Classical method | Supervised fit | 50.7% (45.6%–55.8%) | 0.494 — Mean across held-out participants | 8 | 70 | 1.4 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 10.8% (9.5%–12.2%) | 0.093 — Mean across held-out participants | 8 | 70 | 5.0 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 33.7% (29.8%–37.6%) | 0.318 — Mean across held-out participants | 8 | 70 | 4.9 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 55.8% (50.1%–61.3%) | 0.541 — Mean across held-out participants | 8 | 70 | 285.8 s | | [FBCCA](https://bci.report/methods/fbcca/) — Classical method · FBCCA-M3-N7-H5 · Local Ubuntu / CPU, training-free | Known-frequency reference · no fitting | 79.1% (75.1%–82.8%) | 0.789 — Mean across held-out participants | 8 | 70 | — | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 23.0% (20.3%–25.9%) | 0.200 — Mean across held-out participants | 8 | 70 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 52.8% (47.3%–58.2%) | 0.513 — Mean across held-out participants | 8 | 70 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [CSP+LDA](https://bci.report/methods/csp-lda/). The published CSP+LDA recipe is binary and computes one broadband log-variance per spatial filter over the whole window; the 40 BETA targets lie in one 8–15.8 Hz band, 0.2 Hz apart. **Not run under this protocol:** Temporal ridge, [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/), [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/). ## How these rows were run ### FBCCA on beta-8ch and beta-4ch - Recipe: FBCCA: FBCCA-M3-N7-H5 - Hardware: FBCCA: Local Ubuntu / CPU, training-free - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) Filter-bank CCA, recipe FBCCA-M3-N7-H5: seven Chebyshev sub-bands (6–90 Hz), five harmonics, weights n^−1.25 + 0.25 (Chen et al. 2015, M3); our SciPy realisation. It reads the authors' 3–100 Hz window, and its own filter bank takes the place of the 6–80 Hz step. Window, electrodes, targets and people are those of the matrix. No training; folds unused. FBCCA and CCA are given the 40 stimulus frequencies and use no training data. The same recipe appears on card 11 (other endpoints) and in the electrode-shift study (the 8-electrode figure); these rows give its 40-target value under the core protocol, and the 8-electrode figure was known before this run. [New SSVEP frequencies →](https://bci.report/topics/new-ssvep-frequencies/#practical-settings) [Fewer electrodes →](https://bci.report/topics/fewer-electrodes/#cap-shift) - **FBCCA**: FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free ### ShallowFBCSPNet and Deep4Net under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW - Hardware: ShallowFBCSPNet, Deep4Net: Apple M5 / MPS - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### FBCCA on beta-8ch and beta-4ch | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | FBCCA − Standard CCA — FBCCA-M3-N7-H5 | +16.0 pp | 11.6 to 20.5 | 12.5 to 19.6 | higher | The difference belongs to the whole recipe (filter bank, band edges, five against three harmonics). against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | FBCCA − EEGNet — FBCCA-M3-N7-H5 | +23.3 pp | 17.6 to 29.6 | 18.8 to 28.2 | higher | 98.75 % intervals are Bonferroni-adjusted for four contrasts; the 95 % intervals beside them are unadjusted. Row intervals are the matrix's descriptive 95 % participant bootstrap. ### ShallowFBCSPNet and Deep4Net under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | Deep4Net − EEGNet | −3.0 pp | −3.9 to −2.1 | −4.4 to −1.6 | lower | | ShallowFBCSPNet − EEGNet | −32.7 pp | −37.0 to −28.5 | −38.9 to −26.4 | lower | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 50.71% (45.56%–55.76%) | 0.494 | 8 | 70 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 5.26% (4.75%–5.79%) | 0.034 | 8 | 70 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: −45.45 pp (−50.37 to −40.63): below. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus REVE Large, a frozen-encoder row chosen on these same test folds: −47.48 pp (−52.40 to −42.60): below. ### Notes on this comparator Run as a broad-band control stated before scoring: 5 broad bands cannot resolve 40 targets 0.2 Hz apart. Chance is 2.5%. - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Standard CCA: 63.1% (57.2%–69.0%) - Spectral ridge: 50.7% (45.6%–55.8%) - LaBraM: 10.8% (9.5%–12.2%) - CBraMod: 33.7% (29.8%–37.6%) - EEGNet: 55.8% (50.1%–61.3%) - REVE Base: 37.5% (33.3%–41.9%) - REVE Large: 52.7% (47.7%–57.8%) - LUNA Base: 8.6% (7.7%–9.5%) - LUNA Large: 10.0% (8.9%–11.1%) - BrainOmni Base: 21.7% (18.6%–25.0%) - CodeBrain: 11.9% (10.5%–13.3%) - EEGMamba: 40.2% (35.7%–44.7%) - ST-EEGFormer Base: 17.1% (15.5%–18.8%) - ST-EEGFormer Large: 22.6% (20.1%–25.3%) - eeg-fm-masking MAE (r=9 cm, L=2): 54.8% (49.4%–60.0%) - ERP-FM Base: 18.4% (16.3%–20.5%) - SingLEM: 26.6% (23.7%–29.5%) - ZUNA 1.1: 11.9% (10.8%–13.0%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 37.5% (33.3%–41.9%) | 0.352 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 52.7% (47.7%–57.8%) | 0.513 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 8.6% (7.7%–9.5%) | 0.065 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 10.0% (8.9%–11.1%) | 0.082 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | 21.7% (18.6%–25.0%) | 0.200 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 11.9% (10.5%–13.3%) | 0.100 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 40.2% (35.7%–44.7%) | 0.382 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 17.1% (15.5%–18.8%) | 0.138 | 8 | 70 | in the authors' published pretraining list — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 22.6% (20.1%–25.3%) | 0.190 | 8 | 70 | in the authors' published pretraining list — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 54.8% (49.4%–60.0%) | 0.535 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 18.4% (16.3%–20.5%) | 0.153 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 26.6% (23.7%–29.5%) | 0.197 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 11.9% (10.8%–13.0%) | 0.080 | 8 | 70 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2607.27308) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 54.8% (49.4%–60.0%) | 0.535 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 49.6% (44.5%–54.8%) | 0.481 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 53.4% (48.1%–58.6%) | 0.522 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 51.0% (45.8%–56.1%) | 0.496 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. LUNA min-max scales electrode positions over the channel set; on 4-8 channel montages a small posterior cluster is stretched to the unit cube. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. LUNA min-max position scaling on an 8-channel montage. - **BrainOmni Base**: BrainOmni keeps its upstream common average reference; on 8 channels it removes a median 0.79 of segment channel variance. - **ST-EEGFormer Base**: PRETRAINING-EXPOSED: BETA is in the ST-EEGFormer pretraining corpus (all 70 participants assumed). - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. PRETRAINING-EXPOSED: BETA is in the ST-EEGFormer pretraining corpus. - **ERP-FM Base**: Outside the ERP-FM design: negative control. - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-beta-8ch.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Standard CCA: 63.1% (57.2%–69.0%) - Spectral ridge: 50.7% (45.6%–55.8%) - LaBraM: 10.8% (9.5%–12.2%) - CBraMod: 33.7% (29.8%–37.6%) - EEGNet: 55.8% (50.1%–61.3%) - EEGPT: 20.3% (18.2%–22.6%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 20.3% (18.2%–22.6%) | 0.169 | 8 | 70 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - PO5 and PO6 use channel embeddings that EEGPT’s pretraining never trained: they are two of the four names in its 62-name channel dictionary (AF7, AF8, PO5, PO6) that its 58-electrode pretraining montage leaves out. The other six channels here, and all four of the 4-channel subset, use trained embeddings. - For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated. - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Standard CCA: EEGPT below, the intervals 34.64 pp apart - Spectral ridge: EEGPT below, the intervals 22.96 pp apart - LaBraM: EEGPT above, the intervals 5.97 pp apart - CBraMod: EEGPT below, the intervals 7.24 pp apart - EEGNet: EEGPT below, the intervals 27.54 pp apart EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 27.4% (23.6%–31.5%) | 0.255 | — | | Quark-L — Local Ubuntu / CPU | 27.1% (23.3%–31.0%) | 0.249 | — | | Quark-H — Local Ubuntu / CPU | 23.7% (20.2%–27.3%) | 0.220 | — | | Quark-G † — Local Ubuntu / CUDA | 25.0% (21.3%–28.8%) | 0.234 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 25.8% (22.4%–29.4%) | 0.229 | — | | Molecule-L — Local Ubuntu / CPU | 26.6% (22.9%–30.5%) | 0.241 | — | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 39.3% (35.2%–43.6%) | 0.372 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Limits Near-floor scores are not ordered reliably between the 8- and 4-electrode subsets. Electrode subsets from laboratory recordings do not validate a physical low-channel cap. Prompted SSVEP does not measure idle false activations. Single-seed results; pretraining overlap unknown. On “pretraining overlap unknown”: LaBraM’s and CBraMod’s exposure was checked on 2026-10-04 against their authors’ published pretraining lists. ### The protocol, step by step - Seven participant-disjoint folds: train on 60 people, test on ten. All four blocks stay with their participant. - Two seconds from stimulus onset; no visual-latency shift. Source data were already zero-phase filtered. Additional 6–80 Hz filtering applies to each selected window separately. - Microvolt units are inferred from an independently documented loader, not explicitly stated in the author MAT description. Inconsistent phase metadata are unused by all methods. - Standard CCA uses known frequencies and three harmonics without training labels. Frozen encoders use training-only standardized ridge heads (alpha 100). EEGNet trains from scratch for 20 epochs with one seed (20260912). - Electrodes: PO5, PO3, POZ, PO4, PO6, O1, OZ, O2. - Uniform-guessing reference: 2.5%. Descriptive 95% intervals resample participants; training sets overlap across folds. - No cross-task overall ranking, model fine-tuning optimum or hardware benchmark is claimed. ### Stability One fixed seed and training budget; multi-seed sensitivity pending. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. ### Model terms - **Standard CCA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **FBCCA**: Training-free; no weights. This project's own SciPy code; sub-band design after TRCA-SSVEP filterbank.m (MIT). - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### BETA **Credit** Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. **Licence** [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) [Dataset record ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [Every result on this dataset →](https://bci.report/datasets/beta/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/beta-8ch-protocol.json) **Public-data register note** Eligible only under the stated personal, research-led, noncommercial operation. Any ads, sponsorship, fees, or use directed toward commercial advantage requires fresh review or permission. **Rights reviewed** 2026-09-20 · against [figshare.com ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [doi.org ↗](https://doi.org/10.3389/fnins.2020.00627) ## Reproducibility record - Protocol id: `beta-ssvep-2s-posterior4and8-subject7fold-v1/posterior8` - Dataset release: Hugging Face mirror of BETA database, Figshare record 12264401 v3 · mirror d4290c0200db8a104e0f557349dc49f90ba79506 · upstream Figshare version 3, 2022-06-15 - Hardware: Apple M5 / MPS and CPU - Scoring stage sum: 298.8 s · may include accelerator waiting - Audit record SHA-256: `ccddf911864b8471f488e46cae45a3651369dae1134afcdeaede68bdfcc31950` - Summary SHA-256: `547a35763f7bca10a65c7b44c2f8c46af290159c692d237e92cd682fc21127a6` - Protocol SHA-256: `dcd1d2bf4dbb46b7db8c4562623448633cde4641fa372c90f25ba9d84012c2f0` ### Rows of core-matrix-update-20261106 #### FBCCA on beta-8ch and beta-4ch - Protocol SHA-256: `e4aa044b2cae88f9fe7815c4ab4e83d79ff740b1c6525e046ca86f45cbe7a698` - Audit record SHA-256: `66b5815b8c42bc9383af7db706a67a12fddb21c99448b36668ca74203090458e` - Hardware: Local Ubuntu / CPU, training-free - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws #### ShallowFBCSPNet and Deep4Net under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Apple M5 / MPS - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/beta-8ch-results.csv) - [Protocol · JSON ↓](https://bci.report/data/beta-8ch-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *SSVEP · 8 channels on BETA*. https://bci.report/protocols/beta-8ch/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/beta-8ch/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # SSVEP · 4 channels on BETA Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: 40 visual targets · 4 posterior electrodes - Evaluation: Transfer to a new person - Dataset: [BETA](https://bci.report/datasets/beta/) - People: 70 - Data: 11,200 trials · 7 participant-disjoint folds - Input: 4 channels selected from a 64-channel recording · 2-second windows - Chance level: 2.5% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Standard CCA: 57.6% (51.8%–63.5%) - Spectral ridge: 48.2% (43.1%–53.3%) - LaBraM: 12.9% (11.2%–14.6%) - CBraMod: 27.6% (24.1%–31.3%) - EEGNet: 44.1% (38.6%–49.6%) - FBCCA: 72.3% (67.5%–76.9%) - ShallowFBCSPNet: 16.2% (14.1%–18.6%) - Deep4Net: 38.7% (33.7%–43.8%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | [Standard CCA](https://bci.report/methods/cca/) — Classical method | Known-frequency reference · no fitting | 57.6% (51.8%–63.5%) | 0.567 — Mean across held-out participants | 4 | 70 | 1.1 s | | Spectral ridge — Classical method | Supervised fit | 48.2% (43.1%–53.3%) | 0.467 — Mean across held-out participants | 4 | 70 | 0.7 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 12.9% (11.2%–14.6%) | 0.110 — Mean across held-out participants | 4 | 70 | 2.5 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 27.6% (24.1%–31.3%) | 0.257 — Mean across held-out participants | 4 | 70 | 3.0 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 44.1% (38.6%–49.6%) | 0.421 — Mean across held-out participants | 4 | 70 | 146.7 s | | [FBCCA](https://bci.report/methods/fbcca/) — Classical method · FBCCA-M3-N7-H5 · Local Ubuntu / CPU, training-free | Known-frequency reference · no fitting | 72.3% (67.5%–76.9%) | 0.720 — Mean across held-out participants | 4 | 70 | — | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 16.2% (14.1%–18.6%) | 0.134 — Mean across held-out participants | 4 | 70 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 38.7% (33.7%–43.8%) | 0.373 — Mean across held-out participants | 4 | 70 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [CSP+LDA](https://bci.report/methods/csp-lda/). The published CSP+LDA recipe is binary and computes one broadband log-variance per spatial filter over the whole window; the 40 BETA targets lie in one 8–15.8 Hz band, 0.2 Hz apart. **Not run under this protocol:** Temporal ridge, [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/), [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/). ## How these rows were run ### FBCCA on beta-8ch and beta-4ch - Recipe: FBCCA: FBCCA-M3-N7-H5 - Hardware: FBCCA: Local Ubuntu / CPU, training-free - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) Filter-bank CCA, recipe FBCCA-M3-N7-H5: seven Chebyshev sub-bands (6–90 Hz), five harmonics, weights n^−1.25 + 0.25 (Chen et al. 2015, M3); our SciPy realisation. It reads the authors' 3–100 Hz window, and its own filter bank takes the place of the 6–80 Hz step. Window, electrodes, targets and people are those of the matrix. No training; folds unused. FBCCA and CCA are given the 40 stimulus frequencies and use no training data. The same recipe appears on card 11 (other endpoints) and in the electrode-shift study (the 8-electrode figure); these rows give its 40-target value under the core protocol, and the 8-electrode figure was known before this run. [New SSVEP frequencies →](https://bci.report/topics/new-ssvep-frequencies/#practical-settings) [Fewer electrodes →](https://bci.report/topics/fewer-electrodes/#cap-shift) - **FBCCA**: FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free ### ShallowFBCSPNet and Deep4Net under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW - Hardware: ShallowFBCSPNet, Deep4Net: Apple M5 / MPS - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### FBCCA on beta-8ch and beta-4ch | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | FBCCA − Standard CCA — FBCCA-M3-N7-H5 | +14.7 pp | 10.7 to 19.1 | 11.5 to 18.1 | higher | The difference belongs to the whole recipe (filter bank, band edges, five against three harmonics). against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | FBCCA − EEGNet — FBCCA-M3-N7-H5 | +28.2 pp | 22.8 to 33.9 | 24.0 to 32.6 | higher | 98.75 % intervals are Bonferroni-adjusted for four contrasts; the 95 % intervals beside them are unadjusted. Row intervals are the matrix's descriptive 95 % participant bootstrap. ### ShallowFBCSPNet and Deep4Net under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | Deep4Net − EEGNet | −5.5 pp | −6.8 to −4.1 | −7.5 to −3.5 | lower | | ShallowFBCSPNet − EEGNet | −27.9 pp | −31.9 to −23.9 | −34.4 to −21.9 | lower | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 48.21% (43.06%–53.29%) | 0.467 | 4 | 70 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 5.30% (4.78%–5.84%) | 0.034 | 4 | 70 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: −42.91 pp (−47.74 to −38.10): below. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus REVE Large, a frozen-encoder row chosen on these same test folds: −41.05 pp (−45.48 to −36.69): below. ### Notes on this comparator Run as a broad-band control stated before scoring: 5 broad bands cannot resolve 40 targets 0.2 Hz apart. Chance is 2.5%. - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Standard CCA: 57.6% (51.8%–63.5%) - Spectral ridge: 48.2% (43.1%–53.3%) - LaBraM: 12.9% (11.2%–14.6%) - CBraMod: 27.6% (24.1%–31.3%) - EEGNet: 44.1% (38.6%–49.6%) - REVE Base: 25.5% (22.4%–28.8%) - REVE Large: 46.4% (41.7%–51.2%) - LUNA Base: 9.0% (8.1%–10.1%) - LUNA Large: 11.2% (9.8%–12.6%) - BrainOmni Base: 13.0% (10.6%–15.6%) - CodeBrain: 11.2% (9.8%–12.7%) - EEGMamba: 32.7% (28.9%–36.7%) - ST-EEGFormer Base: 17.5% (15.8%–19.3%) - ST-EEGFormer Large: 22.4% (19.9%–25.0%) - eeg-fm-masking MAE (r=9 cm, L=2): 47.1% (42.1%–52.1%) - ERP-FM Base: 20.9% (18.4%–23.5%) - SingLEM: 20.5% (17.8%–23.3%) - ZUNA 1.1: 8.4% (7.5%–9.4%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 25.5% (22.4%–28.8%) | 0.232 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 46.4% (41.7%–51.2%) | 0.445 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 9.0% (8.1%–10.1%) | 0.069 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 11.2% (9.8%–12.6%) | 0.090 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | 13.0% (10.6%–15.6%) | 0.119 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 11.2% (9.8%–12.7%) | 0.091 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 32.7% (28.9%–36.7%) | 0.305 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 17.5% (15.8%–19.3%) | 0.148 | 4 | 70 | in the authors' published pretraining list — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 22.4% (19.9%–25.0%) | 0.193 | 4 | 70 | in the authors' published pretraining list — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 47.1% (42.1%–52.1%) | 0.454 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 20.9% (18.4%–23.5%) | 0.175 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 20.5% (17.8%–23.3%) | 0.153 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 8.4% (7.5%–9.4%) | 0.053 | 4 | 70 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2607.27308) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 47.1% (42.1%–52.1%) | 0.454 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 43.4% (38.7%–48.3%) | 0.417 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 46.7% (41.9%–51.5%) | 0.450 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 45.4% (40.6%–50.2%) | 0.438 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. LUNA min-max position scaling on a 4-channel montage (see beta-8ch). - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. LUNA min-max position scaling on a 4-channel montage. - **BrainOmni Base**: BrainOmni common average reference on 4 channels removes a median 0.87 of segment channel variance. - **ST-EEGFormer Base**: PRETRAINING-EXPOSED: BETA is in the ST-EEGFormer pretraining corpus. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. PRETRAINING-EXPOSED: BETA is in the ST-EEGFormer pretraining corpus. - **ERP-FM Base**: Outside the ERP-FM design: negative control. - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-beta-4ch.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Standard CCA: 57.6% (51.8%–63.5%) - Spectral ridge: 48.2% (43.1%–53.3%) - LaBraM: 12.9% (11.2%–14.6%) - CBraMod: 27.6% (24.1%–31.3%) - EEGNet: 44.1% (38.6%–49.6%) - EEGPT: 17.3% (15.5%–19.2%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 17.3% (15.5%–19.2%) | 0.144 | 4 | 70 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - For context: BETA is not in EEGPT’s published pretraining list, but the Tsinghua 40-target SSVEP benchmark (Wang et al., 2017) is, as TSU. BETA’s authors describe BETA as an extension of that benchmark, from the same group with the same 40-target layout and its own recordings; whether anyone took part in both is not stated. - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Standard CCA: EEGPT below, the intervals 32.62 pp apart - Spectral ridge: EEGPT below, the intervals 23.89 pp apart - LaBraM: EEGPT above, the intervals 0.90 pp apart - CBraMod: EEGPT below, the intervals 4.91 pp apart - EEGNet: EEGPT below, the intervals 19.47 pp apart EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 27.8% (24.0%–31.9%) | 0.261 | — | | Quark-L — Local Ubuntu / CPU | 28.0% (24.0%–32.2%) | 0.259 | — | | Quark-H — Local Ubuntu / CPU | 26.2% (22.3%–30.3%) | 0.244 | — | | Quark-G † — Local Ubuntu / CUDA | 26.8% (22.8%–31.1%) | 0.255 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 28.0% (24.3%–31.8%) | 0.251 | — | | Molecule-L — Local Ubuntu / CPU | 28.9% (25.1%–33.0%) | 0.265 | — | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 35.9% (32.0%–40.0%) | 0.334 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Limits Near-floor scores are not ordered reliably between the 8- and 4-electrode subsets. Electrode subsets from laboratory recordings do not validate a physical low-channel cap. Prompted SSVEP does not measure idle false activations. Single-seed results; pretraining overlap unknown. On “pretraining overlap unknown”: LaBraM’s and CBraMod’s exposure was checked on 2026-10-04 against their authors’ published pretraining lists. ### The protocol, step by step - Seven participant-disjoint folds: train on 60 people, test on ten. All four blocks stay with their participant. - Two seconds from stimulus onset; no visual-latency shift. Source data were already zero-phase filtered. Additional 6–80 Hz filtering applies to each selected window separately. - Microvolt units are inferred from an independently documented loader, not explicitly stated in the author MAT description. Inconsistent phase metadata are unused by all methods. - Standard CCA uses known frequencies and three harmonics without training labels. Frozen encoders use training-only standardized ridge heads (alpha 100). EEGNet trains from scratch for 20 epochs with one seed (20260912). - Electrodes: POZ, O1, OZ, O2. - Uniform-guessing reference: 2.5%. Descriptive 95% intervals resample participants; training sets overlap across folds. - No cross-task overall ranking, model fine-tuning optimum or hardware benchmark is claimed. ### Stability One fixed seed and training budget; multi-seed sensitivity pending. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. ### Model terms - **Standard CCA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **FBCCA**: Training-free; no weights. This project's own SciPy code; sub-band design after TRCA-SSVEP filterbank.m (MIT). - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### BETA **Credit** Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. **Licence** [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) [Dataset record ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [Every result on this dataset →](https://bci.report/datasets/beta/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/beta-4ch-protocol.json) **Public-data register note** Eligible only under the stated personal, research-led, noncommercial operation. Any ads, sponsorship, fees, or use directed toward commercial advantage requires fresh review or permission. **Rights reviewed** 2026-09-20 · against [figshare.com ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) · [doi.org ↗](https://doi.org/10.3389/fnins.2020.00627) ## Reproducibility record - Protocol id: `beta-ssvep-2s-posterior4and8-subject7fold-v1/posterior4` - Dataset release: Hugging Face mirror of BETA database, Figshare record 12264401 v3 · mirror d4290c0200db8a104e0f557349dc49f90ba79506 · upstream Figshare version 3, 2022-06-15 - Hardware: Apple M5 / MPS and CPU - Scoring stage sum: 153.9 s · may include accelerator waiting - Audit record SHA-256: `ccddf911864b8471f488e46cae45a3651369dae1134afcdeaede68bdfcc31950` - Summary SHA-256: `547a35763f7bca10a65c7b44c2f8c46af290159c692d237e92cd682fc21127a6` - Protocol SHA-256: `dcd1d2bf4dbb46b7db8c4562623448633cde4641fa372c90f25ba9d84012c2f0` ### Rows of core-matrix-update-20261106 #### FBCCA on beta-8ch and beta-4ch - Protocol SHA-256: `e4aa044b2cae88f9fe7815c4ab4e83d79ff740b1c6525e046ca86f45cbe7a698` - Audit record SHA-256: `66b5815b8c42bc9383af7db706a67a12fddb21c99448b36668ca74203090458e` - Hardware: Local Ubuntu / CPU, training-free - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws #### ShallowFBCSPNet and Deep4Net under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Apple M5 / MPS - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/beta-4ch-results.csv) - [Protocol · JSON ↓](https://bci.report/data/beta-4ch-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *SSVEP · 4 channels on BETA*. https://bci.report/protocols/beta-4ch/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/beta-4ch/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Arithmetic & rest on EEGMAT Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: Serial subtraction versus resting EEG - Evaluation: Transfer to a new person - Dataset: [EEGMAT](https://bci.report/datasets/eegmat/) - People: 36 - Data: 2,160 epochs · 5 participant-disjoint folds - Input: 19 channels · 2-second windows - Chance level: 50.0% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Spectral ridge: 56.8% (54.6%–58.9%) - LaBraM: 64.6% (60.5%–68.6%) - CBraMod: 62.3% (58.1%–66.5%) - EEGNet: 67.6% (62.8%–72.5%) - TS+LR: 65.5% (59.6%–71.8%) - MDM: 71.5% (65.5%–77.8%) - ShallowFBCSPNet: 67.9% (63.4%–72.4%) - Deep4Net: 61.3% (57.5%–65.2%) - CSP+LDA: 68.3% (62.0%–74.3%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method | Supervised fit | 56.8% (54.6%–58.9%) | 0.558 — Mean across held-out participants | 19 | 36 | 0.6 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 64.6% (60.5%–68.6%) | 0.620 — Mean across held-out participants | 19 | 36 | 2.1 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 62.3% (58.1%–66.5%) | 0.602 — Mean across held-out participants | 19 | 36 | 6.1 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 67.6% (62.8%–72.5%) Single seed — the highest of the seeds run (see Stability) | 0.637 — Mean across held-out participants | 19 | 36 | 117.7 s | | [TS+LR](https://bci.report/methods/riemannian/) — Classical method · OAS · 4–40 Hz · primary row · Local Ubuntu / CPU | Supervised fit | 65.5% (59.6%–71.8%) | 0.592 — Mean across held-out participants | 19 | 36 | — | | [MDM](https://bci.report/methods/riemannian/) — Classical method · OAS · 4–40 Hz · secondary row · Local Ubuntu / CPU | Supervised fit | 71.5% (65.5%–77.8%) | 0.661 — Mean across held-out participants | 19 | 36 | — | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 67.9% (63.4%–72.4%) | 0.644 — Mean across held-out participants | 19 | 36 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Apple M5 / MPS | Scratch · 20 epochs | 61.3% (57.5%–65.2%) | 0.551 — Mean across held-out participants | 19 | 36 | — | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method · broadband CSP · shrinkage LDA · Apple M5 / CPU | Supervised fit | 68.3% (62.0%–74.3%) | 0.649 — Mean across held-out participants | 19 | 36 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [Standard CCA](https://bci.report/methods/cca/), [FBCCA](https://bci.report/methods/fbcca/) (they need the stimulation frequencies, and this protocol has none); Temporal ridge, [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/) (they model the waveform of a stimulus-locked transient potential, and this protocol's classes are not defined by one). [A frozen NeuroLM pilot with its own preprocessing is on the pilot studies page →](https://bci.report/topics/pilot-studies/#neurolm-eegmat) Frozen-encoder probes with their own preprocessing are on the pilot studies page: [NeuroRVQ on three tasks →](https://bci.report/topics/pilot-studies/#neurorvq) · [EEG-Captioner and AXON on ten-second windows →](https://bci.report/topics/pilot-studies/#eegmat-ten-second) ## How these rows were run ### Riemannian and xDAWN baselines - Recipe: TS+LR: OAS · 4–40 Hz · primary row; MDM: OAS · 4–40 Hz · secondary row - Hardware: TS+LR, MDM: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) One fixed configuration per method: library or MOABB-reference settings with the stated deviations. The 4–40 Hz band-pass (or demeaning plus 50 Hz decimation on the ERP protocols) is part of the recipe and is applied to each window on its own. The EEGMAT signals were ICA-cleaned and may be rank-deficient; OAS shrinkage then dominates the low-variance directions of the covariances that TS+LR and MDM read. - **TS+LR**: Primary row · one fixed configuration (MOABB-reference settings) - **MDM**: Secondary row · one fixed configuration (MOABB-reference settings) ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW; CSP+LDA: broadband CSP · shrinkage LDA - Hardware: ShallowFBCSPNet, Deep4Net: Apple M5 / MPS; CSP+LDA: Apple M5 / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) - **CSP+LDA**: Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### Riemannian and xDAWN baselines | Difference | Mean (pp) | 98.75% interval, Bonferroni-adjusted for four contrasts | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | TS+LR − Spectral ridge — TS+LR · OAS · 4–40 Hz | +8.8 pp | 1.6 to 16.1 | 3.0 to 14.6 | higher | Paired participant bootstrap (10,000 draws). The four contrasts against the Spectral ridge cells carry 98.75 % intervals (Bonferroni over four); the other contrasts carry nominal 95 % intervals with no familywise claim. 'Higher' or 'lower' means the interval lies wholly above or below zero; 'unresolved' means it includes zero. ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | CSP+LDA − EEGNet | +0.7 pp | −4.3 to 5.8 | −7.1 to 8.7 | unresolved | | Deep4Net − EEGNet | −6.3 pp | −10.4 to −2.3 | −12.8 to 0.1 | lower | | ShallowFBCSPNet − EEGNet | +0.3 pp | −2.6 to 3.3 | −4.0 to 5.0 | unresolved | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 56.76% (54.58%–58.94%) | 0.558 | 19 | 36 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 64.95% (60.19%–69.72%) | 0.610 | 19 | 36 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: +8.19 pp (+3.98 to +12.59): above. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus REVE Large, the highest of the 14 frozen-encoder rows this comparison was specified against, chosen on these same test folds: −4.07 pp (−8.70 to +0.74): not separated. Choosing it on the test folds leans the comparison against this row. ### Notes on this comparator The band-power rows in the pilot studies panels on this cohort come from their own input chains and are read on their own. Recording order on this protocol: [On these windows and folds this is the same comparator as the third block of the panel on the components of Neural State Prediction on the pilot studies page →](https://bci.report/topics/pilot-studies/#nsp-components-a) - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 56.8% (54.6%–58.9%) - LaBraM: 64.6% (60.5%–68.6%) - CBraMod: 62.3% (58.1%–66.5%) - EEGNet: 67.6% (62.8%–72.5%) - REVE Base: 68.8% (64.1%–73.2%) - REVE Large: 69.0% (63.3%–74.7%) - LUNA Base: 60.7% (56.9%–64.4%) - LUNA Large: 57.4% (54.3%–60.6%) - BrainOmni Base: 61.6% (58.4%–64.9%) - CodeBrain: 63.0% (59.5%–66.5%) - EEGMamba: 60.3% (57.0%–63.4%) - ST-EEGFormer Base: 64.5% (59.0%–70.0%) - ST-EEGFormer Large: 63.4% (57.8%–69.0%) - eeg-fm-masking MAE (r=9 cm, L=2): 63.6% (59.1%–68.1%) - ERP-FM Base: 61.1% (57.3%–65.1%) - SingLEM: 51.5% (49.2%–53.8%) - ZUNA 1.1: 53.3% (49.5%–57.1%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 68.8% (64.1%–73.2%) | 0.661 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 69.0% (63.3%–74.7%) | 0.661 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 60.7% (56.9%–64.4%) | 0.591 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 57.4% (54.3%–60.6%) | 0.561 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | 61.6% (58.4%–64.9%) | 0.603 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 63.0% (59.5%–66.5%) | 0.609 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 60.3% (57.0%–63.4%) | 0.591 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 64.5% (59.0%–70.0%) | 0.616 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 63.4% (57.8%–69.0%) | 0.593 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 63.6% (59.1%–68.1%) | 0.610 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 61.1% (57.3%–65.1%) | 0.580 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 51.5% (49.2%–53.8%) Interval reaches chance level | 0.514 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 53.3% (49.5%–57.1%) Interval reaches chance level | 0.510 | 19 | 36 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2607.27308) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 63.6% (59.1%–68.1%) | 0.610 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 61.6% (56.4%–66.5%) | 0.588 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 64.3% (60.2%–68.2%) | 0.623 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 62.5% (58.5%–66.2%) | 0.601 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Adapted on this protocol: nine encoders, one fixed recipe The published LaBraM adaptation recipe, unchanged: a head trained on the frozen encoder against rank-4 LoRA with the same head, trained for five epochs on the same participant-disjoint folds, over three seeds. The frozen ridge readout in the table above, the trained head and LoRA order the models differently here. 36 people, 5 folds. LoRA adds 32,768 to 428,032 trainable parameters, depending on the model. | Encoder | Frozen encoder + trained head | LoRA rank 4 + head | LoRA minus frozen + head | Trainable parameters (head · LoRA arm) | | --- | --- | --- | --- | --- | | [LaBraM](https://bci.report/methods/labram/) — LaBraM, already published | 56.6% (54.2%–59.0%) | 64.1% (60.2%–67.8%) | +7.5 pp (+4.8 to +10.1 pp) — 30 improved · 6 worse · 0 unchanged The interval excludes zero. | 402 · 38,802 | | [REVE Base](https://bci.report/methods/reve/) | 64.0% (60.4%–67.7%) | 66.3% (61.9%–70.7%) | +2.3 pp (+0.1 to +4.6 pp) — 21 improved · 13 worse · 2 unchanged The interval excludes zero. | 1,026 · 181,250 | | [REVE Large](https://bci.report/methods/reve/) | 65.6% (61.1%–70.1%) | 69.2% (63.4%–74.9%) | +3.6 pp (+0.5 to +6.8 pp) — 21 improved · 13 worse · 2 unchanged The interval excludes zero. | 2,434 · 430,466 | | [LUNA Base](https://bci.report/methods/luna/) | 58.2% (54.5%–61.7%) | 60.0% (55.9%–63.9%) | +1.8 pp (+0.1 to +3.3 pp) — 23 improved · 11 worse · 2 unchanged The interval excludes zero. | 514 · 37,378 | | [BrainOmni Base](https://bci.report/methods/brainomni/) | 64.5% (59.6%–69.2%) | 64.8% (59.6%–69.7%) | +0.3 pp (−1.4 to +1.8 pp) — 18 improved · 14 worse · 4 unchanged The interval includes zero: no change is established. | 16,386 · 106,498 | | [CodeBrain](https://bci.report/methods/codebrain/) | 66.5% (62.3%–70.6%) | 67.1% (62.6%–71.6%) | +0.6 pp (−0.7 to +1.9 pp) — 18 improved · 14 worse · 4 unchanged The interval includes zero: no change is established. | 7,602 · 58,802 | | [EEGMamba](https://bci.report/methods/eegmamba/) | 60.4% (55.7%–65.2%) | 60.8% (56.3%–65.3%) | +0.4 pp (−0.7 to +1.5 pp) — 18 improved · 14 worse · 4 unchanged The interval includes zero: no change is established. | 7,602 · 90,930 | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) | 63.2% (59.7%–66.7%) | 65.1% (61.0%–69.4%) | +1.9 pp (−0.2 to +4.1 pp) — 20 improved · 16 worse · 0 unchanged The interval includes zero: no change is established. | 1,538 · 148,994 | | [SingLEM](https://bci.report/methods/singlem/) | 50.9% (49.5%–52.3%) | 60.6% (56.7%–64.4%) | +9.7 pp (+6.1 to +13.3 pp) — 30 improved · 6 worse · 0 unchanged The interval excludes zero. | 1,218 · 33,986 | | [ZUNA 1.1](https://bci.report/methods/zuna/) Exposure unknown — Model card: research use only, not for diagnosis or clinical use. | 51.2% (49.8%–52.7%) | 56.9% (53.7%–60.1%) | +5.7 pp (+2.5 to +8.7 pp) — 28 improved · 7 worse · 1 unchanged The interval excludes zero. | 1,218 · 296,130 | **Not adapted:** LUNA Large: not run by design: LUNA Large is frozen probes only in this evaluation’s specification; ST-EEGFormer Large: not run by design: ST-EEGFormer Large is frozen probes only; eeg-fm-masking MAE (r=9 cm, L=2): not run by design: eeg-fm-masking is a frozen-probe masking ablation; ERP-FM Base: not run by design: ERP-FM is frozen probes only and EEGMAT arithmetic is outside the ERP design; eeg-fm-masking JEPA (r=9 cm, L=2): not run by design: eeg-fm-masking is a frozen-probe masking ablation; eeg-fm-masking MAE (r=one channel, L=1): not run by design: eeg-fm-masking is a frozen-probe masking ablation; eeg-fm-masking JEPA (r=one channel, L=1): not run by design: eeg-fm-masking is a frozen-probe masking ablation ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - **ERP-FM Base**: Outside the ERP-FM design: negative control. - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-arithmetic-rest.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 56.8% (54.6%–58.9%) - LaBraM: 64.6% (60.5%–68.6%) - CBraMod: 62.3% (58.1%–66.5%) - EEGNet: 67.6% (62.8%–72.5%) - EEGPT: 65.7% (60.9%–70.3%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 65.7% (60.9%–70.3%) | 0.637 | 19 | 36 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Spectral ridge: EEGPT above, the intervals 1.99 pp apart - LaBraM: the intervals overlap - CBraMod: the intervals overlap - EEGNet: the intervals overlap EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 66.0% (62.0%–70.0%) | 0.635 | — | | Quark-L — Local Ubuntu / CPU | 63.8% (59.8%–67.8%) | 0.619 | — | | Quark-H — Local Ubuntu / CPU | 64.5% (60.6%–68.3%) | 0.631 | — | | Quark-G † — Local Ubuntu / CUDA | 59.4% (56.9%–61.7%) | 0.586 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 67.5% (63.7%–71.3%) | 0.655 | its pretraining corpus includes this protocol's dataset | | Molecule-L — Local Ubuntu / CPU | 63.9% (59.9%–68.0%) | 0.612 | its pretraining corpus includes this protocol's dataset | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 65.2% (60.5%–70.0%) | 0.625 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) Two more models · this protocol’s people and folds ## Signal-JEPA and EEGSimpleConv Signal-JEPA as a frozen encoder with a ridge head fitted on the training people, and EEGSimpleConv trained from scratch for 20 epochs, both on this protocol’s windows, people and 5 folds. Each is compared with EEGNet on the same people. | Model | Training mode | Input | Balanced accuracy | Macro F1 | Difference from EEGNet | | --- | --- | --- | --- | --- | --- | | Signal-JEPA — Foundation model · 3,456,852 parameters | Frozen encoder + ridge head | 0.5–40 Hz band-pass, 128 Hz | 51.11% (48.84%–53.47%) Interval reaches chance level | 0.508 | −16.53 pp (−21.53 to −11.25) — vs this page’s EEGNet cell: 67.6% — against the matrix's one-seed EEGNet cell; training variance not included | | EEGSimpleConv — Compact model · 187,070 parameters | Trained from scratch, 20 epochs | the matrix’s 250 Hz windows | 61.90% — mean of three seeds | 0.577 | −5.62 pp — vs EEGNet’s three-seed mean · no interval | Each difference is a paired comparison with EEGNet on the same people. ### Notes on Signal-JEPA and EEGSimpleConv Signal-JEPA’s input: each window band-passed 0.5–40 Hz and resampled to 128 Hz, the format it was pretrained on, in microvolts, recorded channels only. The further-encoder panel follows its own input contract, so this row sits apart from it. Signal-JEPA’s difference is paired, on the same people, against this page’s EEGNet cell: 10,000 participant bootstrap draws, no multiplicity correction. EEGNet’s mean over its three seeds here is 67.52%. EEGSimpleConv: the mean of three seeds, at width 64 and the matrix’s 20-epoch budget, fixed before scoring. Its difference is from EEGNet’s mean over the same three seeds and is printed without an interval. Signal-JEPA’s model card names Lee2019 (OpenBMI) as its pretraining data. ### Terms of Signal-JEPA and EEGSimpleConv - Signal-JEPA: weights MIT, as the model card at revision 51232ee0 states; code BSD-3-Clause (braindecode 1.5.1); the weights are not redistributed. [paper ↗](https://arxiv.org/abs/2403.11772) · [model card ↗](https://huggingface.co/braindecode/signal-jepa) - EEGSimpleConv: code BSD-3-Clause (braindecode 1.5.1); the separate author repository’s licence is unclear; trained from scratch, and no trained weights are shared. [paper ↗](https://arxiv.org/abs/2309.07159) - [Signal-JEPA and EEGSimpleConv · JSON ↓](https://bci.report/data/signal-jepa-eegsimpleconv-update.json) - [Release `signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/#signal-jepa-eegsimpleconv-update-20261106) ## Limits Thirty nonoverlapping 2-second windows from each of two conditions per person. Published signals were already cleaned with ICA; task performance groups are not evaluated. Recording order: the PhysioNet record describes each person’s file ending _1 as rest recorded before the task and the file ending _2 as recorded during mental arithmetic, so recording order and elapsed time move with the label in every result on this protocol. [PhysioNet record ↗](https://physionet.org/content/eegmat/1.0.0/) ### The protocol, step by step - 5 participant-disjoint folds. All recordings from a person stay together. Each person contributes to the held-out predictions once. - First 60 s of each recording; 30 contiguous nonoverlapping 2 s windows; exclude A2-A1 ear-difference and ECG; EDF physical volts converted to microvolts; subtract each channel's window mean; no rejection, filtering, resampling, or learned preprocessing. - One fixed seed (20260919); no early stopping or test-based tuning. EEGNet trains for 20 epochs per fold. Frozen encoders use training-only standardized ridge heads (alpha 100). - Labels: pre-task-rest, mental-arithmetic. - Uniform-guessing reference: 50.00%. Scores weight participants equally. Intervals describe participant variation; cross-validation training sets overlap. - The benchmark detects condition (rest versus serial subtraction), not the good/bad count-quality participant grouping. - Only the first documented 60 seconds is retained even though EDF containers are longer. - The source README reports prior ICA artifact removal, so these are not untouched acquisition signals. - Open Data Commons Attribution License v1.0 applies; retain PhysioNet attribution. - EEGNet three-seed mean 67.52%; sample SD 0.14 percentage points; range 67.36–67.64%. Main table retains the original fixed seed; this is not a confidence interval. ### Stability EEGNet three-seed mean 67.52%; sample SD 0.14 percentage points; range 67.36–67.64%. Main table retains the original fixed seed; this is not a confidence interval. Seeds run for EEGNet: 67.64% · 67.55% · 67.36%; mean 67.52% Same participants, folds, preprocessing and 20-epoch budget. Three seeds measure initialization variability, not population uncertainty. Main table retains its preselected seed; no best-seed selection. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. EEGNet three-seed mean 67.52%; sample SD 0.14 percentage points; range 67.36–67.64%. Main table retains the original fixed seed; this is not a confidence interval. ### Model terms - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **TS+LR**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **MDM**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **CSP+LDA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### EEGMAT **Credit** Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. **Licence** [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) [Dataset record ↗](https://physionet.org/content/eegmat/1.0.0/) · [Every result on this dataset →](https://bci.report/datasets/eegmat/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/arithmetic-rest-protocol.json) **Public-data register note** Attribution license and study-specific approval/consent are documented; eligibility assumes aggregate-only output and exclusion of subject-info fields. **Rights reviewed** 2026-09-20 · against [physionet.org ↗](https://physionet.org/content/eegmat/1.0.0/) · [www.mdpi.com ↗](https://www.mdpi.com/2306-5729/4/1/14) · [opendatacommons.org ↗](https://opendatacommons.org/licenses/by/1-0/) **Correction on record · 2026-11-06**: The arithmetic-rest protocol’s limitations do not say that EEGMAT records rest before the task for every participant, as PhysioNet describes it. The protocol page states it beside the other limitations. [Corrections register →](https://bci.report/releases/#corrections) ## Reproducibility record - Protocol id: `parallel-fixed-subject-folds-v1/physionet-eegmat-1.0.0` - Dataset release: PhysioNet EEG During Mental Arithmetic Tasks 1.0.0 · 1.0.0 - Hardware: Apple M5 / MPS - Scoring stage sum: 126.5 s · may include accelerator waiting - Audit record SHA-256: `2d0e4cec78c8d6e5bfc9f5390ba161f4308759e02feffbb8645d6983d9093292` - Summary SHA-256: `ab607e749da1eedcf88441c5b6fd0db55e2ebae0bbd91d6d18e508605eef625b` - Protocol SHA-256: `5d53ccbd5a00763908162f0b700c0a3236e0d4f9d43b08e42ee31a2ef01aefbf` ### Rows of core-matrix-update-20261106 #### Riemannian and xDAWN baselines - Protocol SHA-256: `4310329a2d6dddaeff64a965750b64237b08220fc23717b1f3d0c1a48f56b055` - Audit record SHA-256: `ce1ee2a0345358c613aa89298c1cb90939df15d68d6c48298b99f9daa7f69e82` - Hardware: Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws #### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Apple M5 / MPS · Apple M5 / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/arithmetic-rest-results.csv) - [Protocol · JSON ↓](https://bci.report/data/arithmetic-rest-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Arithmetic & rest on EEGMAT*. https://bci.report/protocols/arithmetic-rest/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/#signal-jepa-eegsimpleconv-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/arithmetic-rest/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # P300 target ERP on ds006593 Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: Visual target versus nontarget events - Evaluation: Transfer to a new person - Dataset: [ds006593](https://bci.report/datasets/ds006593/) - People: 21 - Data: 2,520 epochs · 5 participant-disjoint folds - Input: 19 channels · 1-second windows - Chance level: 50.0% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Spectral ridge: 51.0% (49.6%–52.4%) - Temporal ridge: 53.8% (51.9%–55.6%) - LaBraM: 49.4% (47.7%–51.2%) - CBraMod: 55.0% (53.3%–56.7%) - EEGNet: 53.3% (51.7%–54.9%) - XDAWNCov+TS+LR: 61.9% (58.8%–65.0%) - xDAWN+LDA: 60.4% (57.7%–63.2%) - ShallowFBCSPNet: 49.7% (49.0%–50.4%) - Deep4Net: 50.8% (50.0%–51.6%) - CSP+LDA: 50.0% (49.1%–50.8%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method | Supervised fit | 51.0% (49.6%–52.4%) Interval reaches chance level | 0.505 — Mean across held-out participants | 19 | 21 | 0.4 s | | Temporal ridge — Classical method | Supervised fit | 53.8% (51.9%–55.6%) | 0.520 — Mean across held-out participants | 19 | 21 | 0.5 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 49.4% (47.7%–51.2%) At or below chance level | 0.486 — Mean across held-out participants | 19 | 21 | 10.6 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 55.0% (53.3%–56.7%) | 0.529 — Mean across held-out participants | 19 | 21 | 4.8 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 53.3% (51.7%–54.9%) | 0.500 — Mean across held-out participants | 19 | 21 | 45.6 s | | [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/) — Classical method · 4 filters · 50 Hz · primary row · Local Ubuntu / CPU | Supervised fit | 61.9% (58.8%–65.0%) | 0.615 — Mean across held-out participants | 19 | 21 | — | | [xDAWN+LDA](https://bci.report/methods/xdawn/) — Classical method · 4 filters · 50 Hz · secondary row · Local Ubuntu / CPU | Supervised fit | 60.4% (57.7%–63.2%) | 0.601 — Mean across held-out participants | 19 | 21 | — | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Local Ubuntu / CUDA | Scratch · 20 epochs | 49.7% (49.0%–50.4%) At or below chance level | 0.427 — Mean across held-out participants | 19 | 21 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Local Ubuntu / CUDA | Scratch · 20 epochs | 50.8% (50.0%–51.6%) | 0.409 — Mean across held-out participants | 19 | 21 | — | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method · broadband CSP · shrinkage LDA · Local Ubuntu / CPU | Supervised fit | 50.0% (49.1%–50.8%) At or below chance level | 0.453 — Mean across held-out participants | 19 | 21 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [Standard CCA](https://bci.report/methods/cca/), [FBCCA](https://bci.report/methods/fbcca/) (they need the stimulation frequencies, and this protocol has none). **Not run under this protocol:** [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/). [A frozen NeuroRVQ probe with its own input chain is on the pilot studies page →](https://bci.report/topics/pilot-studies/#neurorvq) ## How these rows were run ### Riemannian and xDAWN baselines - Recipe: XDAWNCov+TS+LR: 4 filters · 50 Hz · primary row; xDAWN+LDA: 4 filters · 50 Hz · secondary row - Hardware: XDAWNCov+TS+LR, xDAWN+LDA: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) One fixed configuration per method: library or MOABB-reference settings with the stated deviations. The 4–40 Hz band-pass (or demeaning plus 50 Hz decimation on the ERP protocols) is part of the recipe and is applied to each window on its own. - **XDAWNCov+TS+LR**: Primary row · one fixed configuration (MOABB-reference settings) - **xDAWN+LDA**: Secondary row · one fixed configuration (MOABB-reference settings) ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW; CSP+LDA: broadband CSP · shrinkage LDA - Hardware: ShallowFBCSPNet, Deep4Net: Local Ubuntu / CUDA; CSP+LDA: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) - **Deep4Net**: Braindecode's automatically scaled Deep4Net for 1-s input (no temporal pooling, 229-sample classifier, 235,752 parameters), not the original Deep4Net configuration. The interval's lower end, 50.04 %, prints as 50.0; it sits 0.04 pp above the 50 % reference. - **CSP+LDA**: Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. On this ERP protocol it reads band power, not the waveform. ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### Riemannian and xDAWN baselines | Difference | Mean (pp) | Interval | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | XDAWNCov+TS+LR − Spectral ridge — XDAWNCov+TS+LR · 4 filters · 50 Hz — Spectral ridge: 51.0% | +11.0 pp | 6.8 to 15.1 — 98.75% interval, Bonferroni-adjusted for four contrasts | 7.7 to 14.2 | higher | | XDAWNCov+TS+LR − Temporal ridge — XDAWNCov+TS+LR · 4 filters · 50 Hz — Temporal ridge: 53.8% | +8.1 pp | 5.3 to 11.1 — nominal 95% interval | — | higher | Paired participant bootstrap (10,000 draws). The four contrasts against the Spectral ridge cells carry 98.75 % intervals (Bonferroni over four); the other contrasts carry nominal 95 % intervals with no familywise claim. 'Higher' or 'lower' means the interval lies wholly above or below zero; 'unresolved' means it includes zero. ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | CSP+LDA − EEGNet | −3.3 pp | −5.4 to −1.6 | −6.6 to −0.8 | lower | | Deep4Net − EEGNet | −2.5 pp | −4.2 to −0.9 | −5.2 to −0.2 | lower | | ShallowFBCSPNet − EEGNet | −3.6 pp | −5.4 to −1.9 | −6.3 to −0.9 | lower | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Re-training on the CUDA card is not bit-exact; re-training the published semantic-target EEGNet cell there moved it by 0.1 pp. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 50.95% (49.56%–52.38%) Interval reaches chance level | 0.505 | 19 | 21 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 50.44% (49.33%–51.59%) Interval reaches chance level | 0.469 | 19 | 21 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: −0.52 pp (−2.42 to +1.43): not separated. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus CBraMod, the highest of the 13 frozen-encoder rows this comparison was specified against, chosen on these same test folds: −4.52 pp (−6.23 to −2.78): below. Choosing it on the test folds leans the comparison against this row. ### Notes on this comparator On these 1-second windows the 1–4 Hz band partly carries slow drift within the window and the slow ERP itself, as the Spectral ridge row’s lowest bins do. - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 51.0% (49.6%–52.4%) - Temporal ridge: 53.8% (51.9%–55.6%) - LaBraM: 49.4% (47.7%–51.2%) - CBraMod: 55.0% (53.3%–56.7%) - EEGNet: 53.3% (51.7%–54.9%) - REVE Base: 50.5% (48.5%–52.3%) - REVE Large: 54.6% (52.7%–56.5%) - LUNA Base: 48.7% (46.3%–51.1%) - LUNA Large: 48.8% (46.9%–50.8%) - CodeBrain: 50.8% (49.3%–52.2%) - EEGMamba: 52.5% (51.4%–53.5%) - ST-EEGFormer Base: 53.6% (52.3%–54.8%) - ST-EEGFormer Large: 52.7% (51.3%–54.2%) - eeg-fm-masking MAE (r=9 cm, L=2): 49.5% (48.1%–51.0%) - ERP-FM Base: 51.2% (49.3%–53.2%) - SingLEM: 51.2% (49.2%–53.1%) - ZUNA 1.1: 51.6% (50.3%–53.0%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 50.5% (48.5%–52.3%) Interval reaches chance level | 0.501 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 54.6% (52.7%–56.5%) | 0.533 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 48.7% (46.3%–51.1%) At or below chance level | 0.480 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 48.8% (46.9%–50.8%) At or below chance level | 0.484 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | **Not run** — BrainTokenizer needs a 512-sample window (2.0 s at 256 Hz); the published segment is 1 s = 256 samples, so upstream unfold() would zero-pad 256 of 512 samples of every window (invented samples). All upstream downstream tasks use >= 2 s windows. | 19 | — | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 50.8% (49.3%–52.2%) Interval reaches chance level | 0.461 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 52.5% (51.4%–53.5%) | 0.492 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 53.6% (52.3%–54.8%) | 0.510 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 52.7% (51.3%–54.2%) | 0.505 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 49.5% (48.1%–51.0%) At or below chance level | 0.467 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 51.2% (49.3%–53.2%) Interval reaches chance level | 0.507 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 51.2% (49.2%–53.1%) Interval reaches chance level | 0.492 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 51.6% (50.3%–53.0%) | 0.511 | 19 | 21 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2602.18478) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 49.5% (48.1%–51.0%) At or below chance level | 0.467 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 51.6% (50.4%–52.7%) | 0.503 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 51.1% (49.2%–53.0%) Interval reaches chance level | 0.495 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 50.0% (48.7%–51.5%) Interval reaches chance level | 0.478 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. Point estimate below chance; reported as is. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. Point estimate below chance; reported as is. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - **ERP-FM Base**: In ERP-FM design (post-onset 1 s windows without baseline; pretraining used baselined, filtered epochs). - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-p300-target.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 51.0% (49.6%–52.4%) - Temporal ridge: 53.8% (51.9%–55.6%) - LaBraM: 49.4% (47.7%–51.2%) - CBraMod: 55.0% (53.3%–56.7%) - EEGNet: 53.3% (51.7%–54.9%) - EEGPT: 54.0% (51.5%–56.6%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 54.0% (51.5%–56.6%) | 0.528 | 19 | 21 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - On the P300 and semantic targets, one-second windows ran as four real 0.25-second patches: the patch grid was set to four after the encoder was built, with no padding. EEGPT’s encoder treats each time patch on its own, so this is numerically the same as building it for 256-sample inputs, as the Orthosis pilot built it for its 128-sample windows. - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Spectral ridge: the intervals overlap - Temporal ridge: the intervals overlap - LaBraM: EEGPT above, the intervals 0.28 pp apart - CBraMod: the intervals overlap - EEGNet: the intervals overlap EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 49.6% (48.5%–50.6%) — interval includes chance | 0.445 | — | | Quark-L — Local Ubuntu / CPU | 51.0% (49.4%–52.6%) — interval includes chance | 0.463 | — | | Quark-H — Local Ubuntu / CPU | 52.6% (51.1%–54.1%) | 0.500 | — | | Quark-G † — Local Ubuntu / CUDA | 50.3% (48.5%–52.1%) — interval includes chance | 0.490 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 53.2% (51.5%–54.9%) | 0.520 | its pretraining corpus includes this protocol's dataset | | Molecule-L — Local Ubuntu / CPU | 51.4% (49.2%–53.5%) — interval includes chance | 0.509 | its pretraining corpus includes this protocol's dataset | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 54.3% (52.9%–55.9%) | 0.534 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Limits Balanced target/nontarget sample changes the source 1:9 prevalence. One-second windows include later flashes at 0.3-second intervals; this is not an online speller estimate. ### The protocol, step by step - 5 participant-disjoint folds. All recordings from a person stay together. Each person contributes to the held-out predictions once. - units: microvolts; epoch: [stimulus onset, onset + 1.0 s); filtering: none added - One fixed seed (20260919); no early stopping or test-based tuning. EEGNet trains for 20 epochs per fold. Frozen encoders use training-only standardized ridge heads (alpha 100). - Labels: nontarget, target. - Uniform-guessing reference: 50.00%. Scores weight participants equally. Intervals describe participant variation; cross-validation training sets overlap. - Research subset with fixed deterministic sampling capped at 60 epochs per participant/class. - The cap balances target and nontarget and therefore changes the original approximately 1:9 class prevalence. - Stimuli were presented about every 0.3 s, so each 1 s epoch contains responses to later stimuli; this is intrinsic next-stimulus contamination. ### Stability One fixed seed and training budget; multi-seed sensitivity pending. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. ### Model terms - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Temporal ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **XDAWNCov+TS+LR**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **xDAWN+LDA**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **CSP+LDA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### ds006593 **Credit** OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. **Licence** [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) [Dataset record ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) · [Every result on this dataset →](https://bci.report/datasets/ds006593/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/p300-target-protocol.json) **Public-data register note** Pinned CC0 release with dataset-specific IRB and consent evidence and comparatively low metadata exposure. **Rights reviewed** 2026-09-20 · against [doi.org ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) · [nemar.org ↗](https://nemar.org/dataset/on006593) ## Reproducibility record - Protocol id: `parallel-fixed-subject-folds-v1/ds006593` - Dataset release: OpenNeuro snapshot 1.0.0 · b3fa345b310abfcf442dd4e2867f6652336c3fe4 - Hardware: Local Ubuntu / CUDA - Scoring stage sum: 61.9 s · may include accelerator waiting - Audit record SHA-256: `30762120c57bede9f47d2e96637656e8948fca458016f7f2e3a39f951a357716` - Summary SHA-256: `a63a7ba5bbd6dd7865ae260b707c2db896136646a910d96516fa5c862e1146c0` - Protocol SHA-256: `bf6796dfe04dffa57bdd08933d2bb743722da71594135f04fa4fc33f3fb7dd46` ### Rows of core-matrix-update-20261106 #### Riemannian and xDAWN baselines - Protocol SHA-256: `4310329a2d6dddaeff64a965750b64237b08220fc23717b1f3d0c1a48f56b055` - Audit record SHA-256: `ce1ee2a0345358c613aa89298c1cb90939df15d68d6c48298b99f9daa7f69e82` - Hardware: Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws #### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Local Ubuntu / CUDA · Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/p300-target-results.csv) - [Protocol · JSON ↓](https://bci.report/data/p300-target-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-p300-target.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *P300 target ERP on ds006593*. https://bci.report/protocols/p300-target/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/p300-target/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Semantic target ERP on TMNRED / ds005383 Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: Reading · target versus nontarget events - Evaluation: Transfer to a new person - Dataset: [TMNRED / ds005383](https://bci.report/datasets/tmnred/) - People: 30 - Data: 3,600 epochs · 5 participant-disjoint folds - Input: 30 channels · 1-second windows - Chance level: 50.0% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Spectral ridge: 51.0% (49.6%–52.5%) - Temporal ridge: 56.3% (54.4%–58.4%) - LaBraM: 53.2% (51.3%–54.9%) - CBraMod: 55.8% (54.2%–57.4%) - EEGNet: 61.4% (58.9%–64.0%) - XDAWNCov+TS+LR: 60.3% (58.2%–62.6%) - xDAWN+LDA: 56.8% (55.1%–58.6%) - ShallowFBCSPNet: 56.7% (54.7%–58.8%) - Deep4Net: 57.4% (55.4%–59.5%) - CSP+LDA: 52.7% (51.4%–54.0%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method | Supervised fit | 51.0% (49.6%–52.5%) Interval reaches chance level | 0.506 — Mean across held-out participants | 30 | 30 | 0.9 s | | Temporal ridge — Classical method | Supervised fit | 56.3% (54.4%–58.4%) | 0.559 — Mean across held-out participants | 30 | 30 | 1.2 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 53.2% (51.3%–54.9%) | 0.519 — Mean across held-out participants | 30 | 30 | 3.4 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 55.8% (54.2%–57.4%) | 0.554 — Mean across held-out participants | 30 | 30 | 23.5 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 20 epochs | 61.4% (58.9%–64.0%) | 0.606 — Mean across held-out participants | 30 | 30 | 83.4 s | | [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/) — Classical method · 4 filters · 50 Hz · primary row · Local Ubuntu / CPU | Supervised fit | 60.3% (58.2%–62.6%) | 0.589 — Mean across held-out participants | 30 | 30 | — | | [xDAWN+LDA](https://bci.report/methods/xdawn/) — Classical method · 4 filters · 50 Hz · secondary row · Local Ubuntu / CPU | Supervised fit | 56.8% (55.1%–58.6%) | 0.562 — Mean across held-out participants | 30 | 30 | — | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Local Ubuntu / CUDA | Scratch · 20 epochs | 56.7% (54.7%–58.8%) | 0.553 — Mean across held-out participants | 30 | 30 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Local Ubuntu / CUDA | Scratch · 20 epochs | 57.4% (55.4%–59.5%) | 0.559 — Mean across held-out participants | 30 | 30 | — | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method · broadband CSP · shrinkage LDA · Local Ubuntu / CPU | Supervised fit | 52.7% (51.4%–54.0%) | 0.493 — Mean across held-out participants | 30 | 30 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [Standard CCA](https://bci.report/methods/cca/), [FBCCA](https://bci.report/methods/fbcca/) (they need the stimulation frequencies, and this protocol has none). **Not run under this protocol:** [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/). ## How these rows were run ### Riemannian and xDAWN baselines - Recipe: XDAWNCov+TS+LR: 4 filters · 50 Hz · primary row; xDAWN+LDA: 4 filters · 50 Hz · secondary row - Hardware: XDAWNCov+TS+LR, xDAWN+LDA: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) One fixed configuration per method: library or MOABB-reference settings with the stated deviations. The 4–40 Hz band-pass (or demeaning plus 50 Hz decimation on the ERP protocols) is part of the recipe and is applied to each window on its own. - **XDAWNCov+TS+LR**: Primary row · one fixed configuration (MOABB-reference settings) - **xDAWN+LDA**: Secondary row · one fixed configuration (MOABB-reference settings) ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW; CSP+LDA: broadband CSP · shrinkage LDA - Hardware: ShallowFBCSPNet, Deep4Net: Local Ubuntu / CUDA; CSP+LDA: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) - **Deep4Net**: Braindecode's automatically scaled Deep4Net for 1-s input (no temporal pooling, 229-sample classifier, 242,627 parameters), not the original Deep4Net configuration. - **CSP+LDA**: Broadband CSP without band-pass filtering, the recipe of the published mi-rest CSP+LDA cell; not a tuned CSP. On this ERP protocol it reads band power, not the waveform. ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### Riemannian and xDAWN baselines | Difference | Mean (pp) | Interval | 95% interval, unadjusted | Reading | | --- | --- | --- | --- | --- | | XDAWNCov+TS+LR − Spectral ridge — XDAWNCov+TS+LR · 4 filters · 50 Hz — Spectral ridge: 51.0% | +9.3 pp | 6.1 to 12.8 — 98.75% interval, Bonferroni-adjusted for four contrasts | 6.7 to 12.0 | higher | | XDAWNCov+TS+LR − Temporal ridge — XDAWNCov+TS+LR · 4 filters · 50 Hz — Temporal ridge: 56.3% | +3.9 pp | 2.0 to 5.9 — nominal 95% interval | — | higher | Paired participant bootstrap (10,000 draws). The four contrasts against the Spectral ridge cells carry 98.75 % intervals (Bonferroni over four); the other contrasts carry nominal 95 % intervals with no familywise claim. 'Higher' or 'lower' means the interval lies wholly above or below zero; 'unresolved' means it includes zero. ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | CSP+LDA − EEGNet | −8.8 pp | −11.1 to −6.4 | −12.2 to −5.2 | lower | | Deep4Net − EEGNet | −4.0 pp | −5.8 to −2.1 | −6.8 to −1.1 | lower | | ShallowFBCSPNet − EEGNet | −4.7 pp | −7.2 to −2.2 | −8.5 to −1.1 | lower | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Re-training on the CUDA card is not bit-exact; re-training the published semantic-target EEGNet cell there moved it by 0.1 pp. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 51.03% (49.56%–52.53%) Interval reaches chance level | 0.506 | 30 | 30 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 56.11% (54.22%–58.00%) | 0.546 | 30 | 30 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: +5.08 pp (+3.06 to +7.06): above. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus REVE Large, the highest of the 13 frozen-encoder rows this comparison was specified against, chosen on these same test folds: −3.56 pp (−5.61 to −1.53): below. Choosing it on the test folds leans the comparison against this row. ### Notes on this comparator On these 1-second windows the 1–4 Hz band partly carries slow drift within the window and the slow ERP itself, as the Spectral ridge row’s lowest bins do. - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 51.0% (49.6%–52.5%) - Temporal ridge: 56.3% (54.4%–58.4%) - LaBraM: 53.2% (51.3%–54.9%) - CBraMod: 55.8% (54.2%–57.4%) - EEGNet: 61.4% (58.9%–64.0%) - REVE Base: 53.7% (51.8%–55.4%) - REVE Large: 59.7% (57.5%–61.8%) - LUNA Base: 56.2% (54.2%–58.2%) - LUNA Large: 57.1% (55.3%–58.9%) - CodeBrain: 54.7% (53.0%–56.6%) - EEGMamba: 54.4% (52.5%–56.3%) - ST-EEGFormer Base: 54.6% (52.9%–56.4%) - ST-EEGFormer Large: 57.8% (56.0%–59.6%) - eeg-fm-masking MAE (r=9 cm, L=2): 54.8% (52.7%–56.9%) - ERP-FM Base: 53.3% (51.7%–54.8%) - SingLEM: 58.5% (56.1%–61.1%) - ZUNA 1.1: 53.9% (52.2%–55.6%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 53.7% (51.8%–55.4%) | 0.531 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 59.7% (57.5%–61.8%) | 0.586 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 56.2% (54.2%–58.2%) | 0.553 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 57.1% (55.3%–58.9%) | 0.566 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | **Not run** — BrainTokenizer needs a 512-sample window (2.0 s at 256 Hz); the published segment is 1 s = 256 samples, so upstream unfold() would zero-pad 256 of 512 samples of every window (invented samples). All upstream downstream tasks use >= 2 s windows. | 30 | — | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 54.7% (53.0%–56.6%) | 0.541 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 54.4% (52.5%–56.3%) | 0.539 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 54.6% (52.9%–56.4%) | 0.540 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 57.8% (56.0%–59.6%) | 0.561 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 54.8% (52.7%–56.9%) | 0.542 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 53.3% (51.7%–54.8%) | 0.527 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 58.5% (56.1%–61.1%) | 0.578 | 30 | 30 | in the authors' published pretraining list — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 53.9% (52.2%–55.6%) | 0.534 | 30 | 30 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2602.18478) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 54.8% (52.7%–56.9%) | 0.542 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 55.8% (53.9%–57.6%) | 0.549 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 54.9% (53.4%–56.4%) | 0.542 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 55.0% (53.4%–56.6%) | 0.543 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - **eeg-fm-masking MAE (r=9 cm, L=2)**: Channel-retained pooling gives 15,360-d features (above the harness ~10k soft guideline). - **ERP-FM Base**: In ERP-FM design (post-onset 1 s windows without baseline). - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. PRETRAINING-EXPOSED: TMNRED (ds005383) is in the SingLEM pretraining corpus. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - **eeg-fm-masking JEPA (r=9 cm, L=2)**: Channel-retained pooling gives 15,360-d features (above the harness ~10k soft guideline). - **eeg-fm-masking MAE (r=one channel, L=1)**: Channel-retained pooling gives 15,360-d features (above the harness ~10k soft guideline). - **eeg-fm-masking JEPA (r=one channel, L=1)**: Channel-retained pooling gives 15,360-d features (above the harness ~10k soft guideline). - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-semantic-target.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 51.0% (49.6%–52.5%) - Temporal ridge: 56.3% (54.4%–58.4%) - LaBraM: 53.2% (51.3%–54.9%) - CBraMod: 55.8% (54.2%–57.4%) - EEGNet: 61.4% (58.9%–64.0%) - EEGPT: 54.6% (53.3%–55.9%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 54.6% (53.3%–55.9%) | 0.540 | 30 | 30 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - On the P300 and semantic targets, one-second windows ran as four real 0.25-second patches: the patch grid was set to four after the encoder was built, with no padding. EEGPT’s encoder treats each time patch on its own, so this is numerically the same as building it for 256-sample inputs, as the Orthosis pilot built it for its 128-sample windows. - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Spectral ridge: EEGPT above, the intervals 0.78 pp apart - Temporal ridge: the intervals overlap - LaBraM: the intervals overlap - CBraMod: the intervals overlap - EEGNet: EEGPT below, the intervals 3.06 pp apart EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 55.4% (53.1%–57.8%) | 0.543 | — | | Quark-L — Local Ubuntu / CPU | 54.8% (53.0%–56.6%) | 0.543 | — | | Quark-H — Local Ubuntu / CPU | 52.2% (50.3%–53.9%) | 0.517 | — | | Quark-G † — Local Ubuntu / CUDA | 51.8% (50.1%–53.4%) | 0.514 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 52.8% (51.3%–54.3%) | 0.517 | its pretraining corpus includes this protocol's dataset | | Molecule-L — Local Ubuntu / CPU | 54.7% (52.9%–56.5%) | 0.542 | its pretraining corpus includes this protocol's dataset | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 56.7% (54.5%–59.0%) | 0.558 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Limits Balanced event subset with collapsed semantic categories. Natural class prevalence is not preserved. Third-party reading passages and participant metadata are excluded from this website. ### The protocol, step by step - 5 participant-disjoint folds. All recordings from a person stay together. Each person contributes to the held-out predictions once. - units: microvolts; epoch: [event onset, onset + 1.0 s); filtering: none added - One fixed seed (20260919); no early stopping or test-based tuning. EEGNet trains for 20 epochs per fold. Frozen encoders use training-only standardized ridge heads (alpha 100). - Labels: nontarget, target. - Uniform-guessing reference: 50.00%. Scores weight participants equally. Intervals describe participant variation; cross-validation training sets overlap. - Research subset with fixed deterministic sampling capped at 60 epochs per participant/class. - Balancing to 60 per class changes the source target/nontarget prevalence. - Events use numbered target/nontarget variants; the benchmark collapses the suffix only because the trial_type prefix explicitly names the semantic class. ### Stability One fixed seed and training budget; multi-seed sensitivity pending. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. ### Model terms - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Temporal ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **XDAWNCov+TS+LR**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **xDAWN+LDA**: Deterministic; no weights. pyRiemann 0.12 and scikit-learn, BSD-3-Clause. - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **CSP+LDA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### TMNRED / ds005383 **Credit** Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0. **Licence** [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) · [Every result on this dataset →](https://bci.report/datasets/tmnred/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. [The full review note is in the protocol JSON ↓](https://bci.report/data/semantic-target-protocol.json) **Public-data register note** Strong study-specific open-sharing evidence. Aggregate metrics avoid redistribution of stimulus text and participant metadata; credit under CC BY. **Rights reviewed** 2026-09-20 · against [doi.org ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) · [www.nature.com ↗](https://www.nature.com/articles/s41597-025-05036-2) ## Reproducibility record - Protocol id: `parallel-fixed-subject-folds-v1/ds005383` - Dataset release: OpenNeuro snapshot 1.0.0 · ad78f3db430e636595b1b1c08417492f3067a25e - Hardware: Local Ubuntu / CUDA - Scoring stage sum: 112.4 s · may include accelerator waiting - Audit record SHA-256: `5920974ae9ca3876ce1be82497c139f0783d2ff47173641e329af976441de4c4` - Summary SHA-256: `c5dca7fc30d07a001b8980a4ea65e55ef292f0891a4d32a12262b8ec88389243` - Protocol SHA-256: `839dfba89cf200aac7a28a7259f45024d006fcd4500170e247fcc1c4e9962c8d` ### Rows of core-matrix-update-20261106 #### Riemannian and xDAWN baselines - Protocol SHA-256: `4310329a2d6dddaeff64a965750b64237b08220fc23717b1f3d0c1a48f56b055` - Audit record SHA-256: `ce1ee2a0345358c613aa89298c1cb90939df15d68d6c48298b99f9daa7f69e82` - Hardware: Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws #### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Local Ubuntu / CUDA · Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/semantic-target-results.csv) - [Protocol · JSON ↓](https://bci.report/data/semantic-target-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-semantic-target.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Sleep staging on EESM19 scalp subset](https://bci.report/protocols/sleep-scalp/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Semantic target ERP on TMNRED / ds005383*. https://bci.report/protocols/semantic-target/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/semantic-target/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Sleep staging on EESM19 scalp subset Core-matrix protocol · Transfer to a new person The released protocol, with the score of every method run under it. ## The protocol at a glance - Task: Five stages · balanced scalp-EEG sample - Evaluation: Transfer to a new person - Dataset: [EESM19 scalp subset](https://bci.report/datasets/eesm19/) - People: 20 - Data: 3,000 epochs · 5 participant-disjoint folds - Input: 6 channels · 30-second windows - Chance level: 20.0% - Metrics: Balanced accuracy (primary) · Macro F1 (secondary) ## Results Compare within this protocol only: other protocols differ in cohort, electrodes, window or chance level. **Balanced accuracy, every method under this protocol.** Dot: the estimate; line: descriptive 95% interval; dashed line: chance level. - Spectral ridge: 72.2% (70.7%–73.8%) - LaBraM: 71.1% (69.0%–73.2%) - CBraMod: 71.3% (68.7%–73.7%) - EEGNet: 56.5% (54.9%–58.3%) - ShallowFBCSPNet: 55.0% (53.1%–56.8%) - Deep4Net: 45.2% (42.7%–47.7%) - CSP+LDA: 50.1% (46.9%–53.1%) | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | Scoring time | | --- | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method | Supervised fit | 72.2% (70.7%–73.8%) | 0.705 — Mean across held-out participants | 6 | 20 | 2.6 s | | [LaBraM](https://bci.report/methods/labram/) — Foundation model | Frozen encoder + ridge head | 71.1% (69.0%–73.2%) | 0.697 — Mean across held-out participants | 6 | 20 | 11.5 s | | [CBraMod](https://bci.report/methods/cbramod/) — Foundation model | Frozen encoder + ridge head | 71.3% (68.7%–73.7%) | 0.702 — Mean across held-out participants | 6 | 20 | 10.2 s | | [EEGNet](https://bci.report/methods/eegnet/) — Compact model | Scratch · 10 epochs | 56.5% (54.9%–58.3%) | 0.545 — Mean across held-out participants | 6 | 20 | 90.4 s | | [ShallowFBCSPNet](https://bci.report/methods/shallowfbcspnet/) — Compact model · matrix recipe · AdamW · Local Ubuntu / CUDA | Scratch · 10 epochs | 55.0% (53.1%–56.8%) | 0.519 — Mean across held-out participants | 6 | 20 | — | | [Deep4Net](https://bci.report/methods/deep4net/) — Compact model · matrix recipe · AdamW · Local Ubuntu / CUDA | Scratch · 10 epochs | 45.2% (42.7%–47.7%) | 0.419 — Mean across held-out participants | 6 | 20 | — | | [CSP+LDA](https://bci.report/methods/csp-lda/) — Classical method · one-vs-rest CSP · shrinkage LDA · Local Ubuntu / CPU | Supervised fit | 50.1% (46.9%–53.1%) | 0.483 — Mean across held-out participants | 6 | 20 | — | Scoring time covers fitting and prediction in this configuration, may include accelerator waiting, and excludes data preparation. Scoring time is not recorded for the rows of core-matrix-update-20261106. **Not applicable under this protocol:** [Standard CCA](https://bci.report/methods/cca/), [FBCCA](https://bci.report/methods/fbcca/) (they need the stimulation frequencies, and this protocol has none); Temporal ridge, [XDAWNCov+TS+LR](https://bci.report/methods/xdawn/), [xDAWN+LDA](https://bci.report/methods/xdawn/) (they model the waveform of a stimulus-locked transient potential, and this protocol's classes are not defined by one). **Not run under this protocol:** [TS+LR](https://bci.report/methods/riemannian/), [MDM](https://bci.report/methods/riemannian/). ## How these rows were run ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Recipe: ShallowFBCSPNet, Deep4Net: matrix recipe · AdamW; CSP+LDA: one-vs-rest CSP · shrinkage LDA - Hardware: ShallowFBCSPNet, Deep4Net: Local Ubuntu / CUDA; CSP+LDA: Local Ubuntu / CPU - Release: [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) - **CSP+LDA**: One-vs-rest CSP: one binary CSP per stage, 30 log-variance features, one shrinkage LDA; broadband as in the published mi-rest recipe; a new code path, not a tuned CSP. ## Paired differences Each difference is the mean over people of each person’s difference in balanced accuracy, in percentage points (pp), on the same windows and people. ### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe against the matrix's one-seed EEGNet cell; training variance not included | Difference | Mean (pp) | 95% interval, unadjusted | Bonferroni interval over 18 contrasts | Reading | | --- | --- | --- | --- | --- | | CSP+LDA − EEGNet — CSP+LDA (one-vs-rest CSP) | −6.4 pp | −9.3 to −3.3 | −10.8 to −1.5 | lower | | Deep4Net − EEGNet | −11.3 pp | −13.6 to −9.0 | −14.7 to −7.8 | lower | | ShallowFBCSPNet − EEGNet | −1.6 pp | −3.4 to 0.4 | −4.7 to 1.5 | unresolved | The paired 95 % interval governs the wording; the Bonferroni interval beside it covers all 18 contrasts. Read at 95 % without correction, about one in 20 contrasts with no true difference may show one by chance. Re-training on the CUDA card is not bit-exact; re-training the published semantic-target EEGNet cell there moved it by 0.1 pp. Beside the core matrix ## Log band-power ridge (5 bands, visible channels) Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. It uses the same windows, people, folds and ridge head as the Spectral ridge row, repeated from the table above. | Method | Training mode | Balanced accuracy | Macro F1 | Channels | People | | --- | --- | --- | --- | --- | --- | | Spectral ridge — Classical method · core matrix row above | Supervised fit | 72.23% (70.70%–73.80%) | 0.705 | 6 | 20 | | Log band-power ridge (5 bands, visible channels) — Classical method | Supervised fit | 66.00% (63.03%–68.93%) | 0.636 | 6 | 20 | ### Differences from two published rows Each difference is this row minus the named row on the same people, with a paired participant bootstrap (10,000 draws): an interval wholly above zero reads above, wholly below reads below, and one that includes zero reads not separated, which does not show the two are equivalent. The intervals are descriptive, conditional on these folds, with no multiplicity correction. Minus the Spectral ridge row: −6.23 pp (−8.50 to −3.90): below. The two rows differ in bands, normalisation, estimator, frequency range and sampling rate at once, so the difference measures the whole recipe. Minus ST-EEGFormer Large, the highest of the 14 frozen-encoder rows this comparison was specified against, chosen on these same test folds: −13.60 pp (−15.80 to −11.37): below. Choosing it on the test folds leans the comparison against this row. - [The comparator’s cells · JSON ↓](https://bci.report/data/logpower-ridge-update.json) - [How it differs from the Spectral ridge row, and two checks of its head and input scaling →](https://bci.report/methods/#log-band-power-ridge) - [Release `logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) Beside the core matrix ## Further foundation encoders, frozen Each encoder replaces LaBraM or CBraMod in the frozen recipe above; the windows, people, folds, ridge head and scoring stay the same. Rows are grouped by model family; a row is above or below another only where their 95% intervals do not overlap. Each interval is a descriptive participant bootstrap with no multiplicity correction, and with 126 frozen cells an occasional non-overlap is expected by chance. Every encoder sees this protocol’s published windows, usually shorter than its pretraining context. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: the further encoders, in family order. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 72.2% (70.7%–73.8%) - LaBraM: 71.1% (69.0%–73.2%) - CBraMod: 71.3% (68.7%–73.7%) - EEGNet: 56.5% (54.9%–58.3%) - REVE Base: 74.6% (72.9%–76.4%) - REVE Large: 76.6% (75.2%–78.0%) - LUNA Base: 67.5% (65.3%–69.5%) - LUNA Large: 72.4% (70.4%–74.0%) - BrainOmni Base: 67.7% (65.8%–69.5%) - CodeBrain: 70.7% (68.8%–72.7%) - EEGMamba: 64.1% (61.6%–66.4%) - ST-EEGFormer Base: 77.3% (75.6%–78.8%) - ST-EEGFormer Large: 79.6% (78.2%–81.0%) - eeg-fm-masking MAE (r=9 cm, L=2): 67.9% (65.4%–70.4%) - ERP-FM Base: 71.0% (69.5%–72.6%) - SingLEM: 35.3% (33.7%–37.0%) - ZUNA 1.1: 60.1% (57.8%–62.3%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [REVE Base](https://bci.report/methods/reve/) 1 — Foundation model | Frozen encoder + ridge head | 74.6% (72.9%–76.4%) | 0.733 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [REVE Large](https://bci.report/methods/reve/) 2 — Foundation model | Frozen encoder + ridge head | 76.6% (75.2%–78.0%) | 0.760 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.21585) | | [LUNA Base](https://bci.report/methods/luna/) 3 — Foundation model | Frozen encoder + ridge head | 67.5% (65.3%–69.5%) | 0.656 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [LUNA Large](https://bci.report/methods/luna/) 4 — Foundation model | Frozen encoder + ridge head | 72.4% (70.4%–74.0%) | 0.707 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2510.22257) | | [BrainOmni Base](https://bci.report/methods/brainomni/) 5 — Foundation model | Frozen encoder + ridge head | 67.7% (65.8%–69.5%) | 0.674 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2505.18185v3) | | [CodeBrain](https://bci.report/methods/codebrain/) 6 — Foundation model | Frozen encoder + ridge head | 70.7% (68.8%–72.7%) | 0.694 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2506.09110) | | [EEGMamba](https://bci.report/methods/eegmamba/) 7 — Foundation model | Frozen encoder + ridge head | 64.1% (61.6%–66.4%) | 0.625 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://github.com/wjq-learning/EEGMamba/blob/main/pretrain_main.py) | | [ST-EEGFormer Base](https://bci.report/methods/st-eegformer/) 8 — Foundation model | Frozen encoder + ridge head | 77.3% (75.6%–78.8%) | 0.765 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) 9 — Foundation model | Frozen encoder + ridge head | 79.6% (78.2%–81.0%) | 0.789 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh) | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model | Frozen encoder + ridge head | 67.9% (65.4%–70.4%) | 0.668 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [ERP-FM Base](https://bci.report/methods/erp-fm/) 11 — Foundation model | Frozen encoder + ridge head | 71.0% (69.5%–72.6%) | 0.695 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.32796) | | [SingLEM](https://bci.report/methods/singlem/) 12 — Foundation model | Frozen encoder + ridge head | 35.3% (33.7%–37.0%) | 0.325 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2509.17920v3) | | [ZUNA 1.1](https://bci.report/methods/zuna/) 13 — Foundation model — Model card: research use only, not for diagnosis or clinical use. | Frozen encoder + ridge head | 60.1% (57.8%–62.3%) | 0.579 | 6 | 20 | Exposure unknown — unknown: the authors do not list their pretraining data closely enough to decide — [source ↗](https://arxiv.org/abs/2602.18478) | Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 2026-10-04: a dataset absent from it is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen. ST-EEGFormer lists BETA and SingLEM lists TMNRED; ZUNA 1.1 publishes no list, so every one of its cells is unknown; EEGMamba’s list is read from its official code (its paper was not read), with medium confidence. ### Row footnotes - Input removes the per-segment per-channel mean before the published /100 scaling (declared before scoring); a no-mean-removal sensitivity run scores P300 0.71 percentage points higher and sleep 1.40 lower. - Input removes the per-segment per-channel mean (declared before scoring); the no-mean-removal sensitivity run scores P300 2.06 percentage points lower and sleep 0.93 lower. - Resampled to 256 Hz and truncated to whole 40-sample patches; 1-2 s windows and 4-8 channel montages are outside the authors' tested regime. - As LUNA Base. Frozen probes only. - p300-target and semantic-target not run: the 2 s tokenizer window cannot be met by 1 s segments without zero padding. Upstream common average reference kept on 4-8 channel montages. - Published CBraMod pooling (patch mean, channels kept); LoRA effectively adapts only the V projection (1-token attention window upstream). - Published CBraMod pooling; LoRA on Mamba2 in_proj and out_proj (24 targets). - BETA cells are pretraining-exposed (BETA is in the ST-EEGFormer corpus). - BETA cells are pretraining-exposed. The authors' downstream class (1-based temporal index) was used; Large may have been pretrained with a 0-based index. Frozen probes only. - Paper-recommended masking geometry; channel-retained pooling (C x 512). Three sibling checkpoints appear in the masking ablation panel. - In design: p300-target and semantic-target. The other six protocols are negative controls outside the model's design. ERP windows have no pre-stimulus baseline. - semantic-target is pretraining-exposed (TMNRED is in the SingLEM corpus). Each channel is encoded independently; the last 0.25 s of each 2 s segment is not seen by the tokenizer. - Pretraining exposure unknown for every core dataset (no published source list). Encoder only; no channel upsampling or reconstruction. - A sibling of the paper-recommended checkpoint in the masking ablation: the same architecture, corpus and recipe, another masking framework or geometry. Masking ablation: three sibling checkpoints of eeg-fm-masking Intervals are marginal: the same people under two checkpoints are not tested as a pair. | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [eeg-fm-masking MAE (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 10 — Foundation model · matrix row above | Frozen encoder + ridge head | 67.9% (65.4%–70.4%) | 0.668 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=9 cm, L=2)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 70.3% (68.0%–72.7%) | 0.695 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking MAE (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 65.2% (63.0%–67.3%) | 0.644 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | | [eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) 14 — Foundation model | Frozen encoder + ridge head | 68.3% (65.7%–70.8%) | 0.675 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-04) — [source ↗](https://arxiv.org/abs/2609.33487) | ### Weights licences and terms - **REVE Base · REVE Large**: REVE Responsible Use License v1.0 (owner-approved 2026-10-04) REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. Model versions: REVE Base @ dc2a075c · REVE Large @ 317531c7. [paper ↗](https://arxiv.org/abs/2510.21585) - **LUNA Base · LUNA Large**: CC BY-ND 4.0 (owner-approved: internal adaptation, aggregate scores only, no adapted weights or LoRA deltas shared) CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. [paper ↗](https://arxiv.org/abs/2510.22257) - **BrainOmni Base**: MIT [paper ↗](https://arxiv.org/abs/2505.18185) - **CodeBrain**: Apache-2.0 [paper ↗](https://arxiv.org/abs/2506.09110) - **EEGMamba**: MIT [paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) - **ST-EEGFormer Base · ST-EEGFormer Large**: MIT (repo LICENSE; README reserves the paper, diagrams and the name) MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). [paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) - **eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)**: CC-BY-4.0 (cite arXiv:2609.33487) CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). [paper ↗](https://arxiv.org/abs/2609.33487) - **ERP-FM Base**: CC BY-NC-SA 4.0 (non-commercial; aggregate scores only) CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. [paper ↗](https://arxiv.org/abs/2609.32796) - **SingLEM**: MIT (checkpoint committed in the MIT repository) MIT. [paper ↗](https://arxiv.org/abs/2509.17920) - **ZUNA 1.1**: Apache-2.0; model card: research use only, not for diagnosis or clinical use Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. [paper ↗](https://arxiv.org/abs/2607.27308) No model’s authors endorse these results, and no adapted weights or LoRA deltas were kept or shared. ### Notes on these rows - **REVE Base**: REVE input removes the per-segment per-channel mean (declared before scoring; the published LaBraM/CBraMod rows did not); a no-mean-removal sensitivity run is stated in the row footnote. - **REVE Large**: REVE input removes the per-segment per-channel mean (declared before scoring); a no-mean-removal sensitivity run is stated in the row footnote. - **LUNA Base**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. LUNA min-max position scaling on a 6-channel montage. - **LUNA Large**: LUNA resamples to 256 Hz and keeps whole 40-sample patches: 1 s keeps 240/256 samples, 2 s keeps 480/512. LUNA min-max position scaling on a 6-channel montage. - **BrainOmni Base**: BrainOmni common average reference on 6 channels removes a median 0.45 of segment channel variance. - **ST-EEGFormer Large**: ST-EEGFormer Large uses the authors' downstream class (1-based temporal index); a label-free reconstruction check suggests Large was pretrained with a 0-based index, so its temporal sinusoid may be shifted by one position. - **ERP-FM Base**: Outside the ERP-FM design: negative control. - **SingLEM**: SingLEM tokenizer (128-sample tokens, stride 96 at 128 Hz) leaves the last 0.25 s of every 2 s segment unseen. - **ZUNA 1.1**: ZUNA pretraining sources are not published: exposure unknown on every core dataset. 0-4 electrode coordinates per protocol are clamped to the edge of the +-0.12 m position grid as upstream does. - [Foundation-encoder rows · CSV ↓](https://bci.report/data/foundation-models-sleep-scalp.csv) - [Release `foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) Beside the core matrix ## EEGPT, frozen EEGPT ran in the same runs as the published LaBraM and CBraMod rows above, with only the encoder swapped: the same windows, people, folds, ridge head and scoring. **Balanced accuracy under this protocol.** Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level. - Spectral ridge: 72.2% (70.7%–73.8%) - LaBraM: 71.1% (69.0%–73.2%) - CBraMod: 71.3% (68.7%–73.7%) - EEGNet: 56.5% (54.9%–58.3%) - EEGPT: 75.5% (73.9%–77.1%) | Encoder | Training mode | Balanced accuracy | Macro F1 | Channels | People | Pretraining exposure | | --- | --- | --- | --- | --- | --- | --- | | [EEGPT](https://bci.report/methods/eegpt/) — Foundation model | Frozen encoder + ridge head | 75.5% (73.9%–77.1%) | 0.748 | 6 | 20 | — not in the authors' published pretraining list (checked 2026-10-12) — [source ↗](https://papers.neurips.cc/paper_files/paper/2024/file/4540d267eeec4e5dbd9dae9448f0b739-Paper-Conference.pdf) | ### Footnotes to this row - Every core window is shorter than EEGPT’s 4-second pretraining window: 1 or 2 seconds, and sleep as fifteen 2-second segments. The encoder computes each 0.25-second patch the same way at any length; a shorter window gives fewer patches to average. - The readout is the site’s fixed probe on the frozen summary tokens, as for the published LaBraM and CBraMod rows: ridge (alpha 100) on 2,048 pooled values, and on idle a per-person logistic head. It is not the authors’ downstream probe, a learnable spatial filter with a linear head, and the windows are fed without the average reference EEGPT was pretrained with. ### Against the published rows Each row of the table above that experiments.json holds, by whether its 95% interval and EEGPT’s overlap (no multiplicity correction): - Spectral ridge: EEGPT above, the intervals 0.10 pp apart - LaBraM: EEGPT above, the intervals 0.67 pp apart - CBraMod: EEGPT above, the intervals 0.17 pp apart - EEGNet: EEGPT above, the intervals 15.60 pp apart EEGPT’s exposure is read from its authors’ published pretraining list, checked on 2026-10-12, as for the further encoders. Its pretraining corpus: Five public datasets, as the paper’s Table 1 lists them: PhysioMI (PhysioNet motor imagery and execution), HGD, TSU (the Tsinghua 40-target SSVEP benchmark), SEED and M3CV; 58 electrodes, 256 Hz, millivolts, average reference, 4-second windows. ### Licence and terms - **Code licence**: Apache-2.0 (the repository’s source code only) - **Weights licence**: unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights - **Weights**: no original or converted weights are distributed; the hashes identify the checkpoint used and grant nothing - **Rights review**: Published by the owner’s decision, aggregate results only, as the pilot comparison is. The checkpoint’s licence could not be read; no publisher prohibition on independently computed aggregate results was found; no original or converted weights are distributed. - [paper ↗](https://doi.org/10.52202/079017-1239) · [repository ↗](https://github.com/BINE022/EEGPT) - [EEGPT’s cells · JSON ↓](https://bci.report/data/eegpt-core-update.json) - [EEGPT’s page →](https://bci.report/methods/eegpt/) - [Release `eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) Frozen-probe comparison ## LaBraM at Scale and REVE Small, frozen Six LaBraM at Scale encoders and REVE Small’s encoder in the same frozen recipe as the rows above: the same windows, people, folds, head and scoring, with no random-initialization control. Balanced accuracy with its 95% interval, and macro F1, frozen encoders. | Encoder | Balanced accuracy (95% interval) | Macro F1 | Exposure | | --- | --- | --- | --- | | LaBraM at Scale · Quark ladder | | | | | Quark-B — Local Ubuntu / CPU | 72.1% (70.3%–74.0%) | 0.701 | — | | Quark-L — Local Ubuntu / CPU | 70.3% (68.0%–72.5%) | 0.686 | — | | Quark-H — Local Ubuntu / CPU | 72.9% (71.5%–74.2%) | 0.716 | — | | Quark-G † — Local Ubuntu / CUDA | 67.3% (65.6%–68.9%) | 0.670 | — | | LaBraM at Scale · Atom-L and Molecule-L | | | | | Atom-L — Local Ubuntu / CPU | 72.7% (70.6%–74.4%) | 0.710 | — | | Molecule-L — Local Ubuntu / CPU | 71.9% (70.2%–73.4%) | 0.697 | — | | REVE Small · early checkpoint (epoch 15 of 200) | | | | | REVE Small — Local Ubuntu / CPU | 64.5% (62.5%–66.6%) | 0.626 | — | † Quark-G: Run on Local Ubuntu / CUDA; its step is read against a GPU re-run of Quark-H and includes the probe's input width doubling from 800 to 1600 under one fixed penalty (see the topic page). Unless a cell is flagged, its dataset is not in that collection’s published pretraining list (the LaBraM at Scale pretraining manifest at faa387d9, checked 2026-11-05); REVE Small’s is read from the REVE corpus list (checked 2026-10-04). The Quark ladder’s paired differences across protocols, the Quark-G step among them, are on “[Does pretraining help EEG foundation models like LaBraM and CBraMod?](https://bci.report/topics/does-pretraining-help/#fc-ladder)”. Terms: LaBraM at Scale’s code and weights under MIT (Pan-Junyu/LaBraM-at-Scale @ a0b0ce27); REVE Small under the REVE Responsible Use License v1.0 (reve-small @ 3565bdcb), aggregate scientific results only. No weights are distributed. - [Download the JSON](https://bci.report/data/foundation-checkpoints-update.json) - [Release `foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) ## Limits Balanced quality-screened scalp subset, not whole-night deployment prevalence and not ear-EEG. Frozen encoders average fifteen 2-second representations; EEGNet receives 10 epochs. ### The protocol, step by step - 5 participant-disjoint folds. All recordings from a person stay together. Each person contributes to the held-out predictions once. - Retain six named scalp electrodes; exclude mastoids. Drop any epoch with source per-channel missing-value flag in those six electrodes, actual nonfinite values, any channel std<0.01uV, or peak-to-peak>1000uV. Use full30s at200Hz, no further filtering/reference/amplitude transformation. Select up to30 evenly spaced eligible epochs per participant/class across available nights. - One fixed seed (20260919); no early stopping or test-based tuning. EEGNet trains for 10 epochs per fold. Frozen encoders use training-only standardized ridge heads (alpha 100). - Labels: Wake, N1, N2, N3, REM. - Uniform-guessing reference: 20.00%. Scores weight participants equally. Intervals describe participant variation; cross-validation training sets overlap. - Lightweight balanced quality-screened subset; scores do not describe natural sleep-stage prevalence or the entire73780epoch release. - Two original corrupt sessions and boundary epochs were excluded by the uploader. - 200Hz data may have finite replacements despite original missing-value flags; source flags are therefore enforced instead of relying on finite checks alone. - One prediction per30s epoch. Frozen encoders pool15nonoverlapping2s segments; no sequence context across epochs. - All nights from a participant remain together; pretraining overlap unknown. - Quality thresholds fixed before any scores; source artifacts may remain. ### Stability One fixed seed and training budget; multi-seed sensitivity pending. ### Pretraining exposure Unknown unless explicitly documented; no unseen-pretraining claim. **Checked 2026-10-04.** LaBraM: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2405.18765); CBraMod: not in the authors' published pretraining list (checked 2026-10-04) [source ↗](https://arxiv.org/abs/2412.07236). The released sentence above predates this check. ### Notes on the methods - One fixed configuration. Foundation encoders remain frozen; small networks train from scratch. These scores do not establish optimal fine-tuned performance. No individual predictions or participant-level results are distributed. ### Model terms - **Spectral ridge**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **LaBraM**: Code/repository: MIT · Checkpoint: committed in that repository; no separate weight terms - **CBraMod**: Code: MIT · Weights: Apache-2.0 (official model card) - **EEGNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **ShallowFBCSPNet**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **Deep4Net**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. - **CSP+LDA**: Trained from scratch / deterministic reference; no third-party pretrained weights. Braindecode BSD-3-Clause; MNE/scikit-learn BSD where used. ## Source and licence ### EESM19 scalp subset **Credit** Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. **Licence** [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) [Dataset record ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) · [Every result on this dataset →](https://bci.report/datasets/eesm19/) The figures here are aggregate measurements computed by BCI Report under the licence above; the data remain their creators’. **Permitted scope** Personal noncommercial research; aggregate results only **Privacy** Only cohort aggregates are published here: no recording, no participant identifier, no per-person score. According to the 2025 data descriptor, the informed consent form did not mention publication, and before release the GDPR office of Region Midt judged the data fully anonymised: consent covered the study, and the public release rests on that anonymisation judgement. [The full review note is in the protocol JSON ↓](https://bci.report/data/sleep-scalp-protocol.json) **Public-data register note** Clear upstream CC0 and study ethics evidence; use only the pinned scalp-signal derivative and publish aggregates, with provenance disclosed. Consent covered the study, not publication; the public release rests on a GDPR anonymization assessment (amended 2026-09-22). **Rights reviewed** 2026-09-20 · against [doi.org ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) · [www.nature.com ↗](https://www.nature.com/articles/s41598-019-53115-3) · [doi.org ↗](https://doi.org/10.1038/s41597-025-04579-8) ## Reproducibility record - Protocol id: `parallel-fixed-subject-folds-v1/eesm19-scalp-sleep` - Dataset release: Hugging Face processed mirror of OpenNeuro ds005185 1.0.2 · mirror fb045012b58a0b755f9fedf5931dff9b64a9797f · upstream 0857858f7a2ba1582930f23eca3ec56f90a96da9 - Hardware: Local Ubuntu / CUDA - Scoring stage sum: 114.7 s · may include accelerator waiting - Audit record SHA-256: `e6cbf1b199bf2902b9c46e4575ebfa8aff47952625df3d71a1d720b11cc27291` - Summary SHA-256: `b7c784c17d6b33c16ffca5df42d3fd8f3d7bf14960675a7278089c4903886567` - Protocol SHA-256: `dace8a98e480ae7ee890dd6f498d5753a90e1573813b6e174f3c14bcbae8a8ff` ### Rows of core-matrix-update-20261106 #### ShallowFBCSPNet, Deep4Net and CSP+LDA under the matrix recipe - Protocol SHA-256: `a216bfac4ce6dd11c63d499b58bc93aebb89025168911b5d7d7fe265a866da4c` - Audit record SHA-256: `f5a1c3dab4a7e265fa3413ada48edc89b9fd4ac98ab7dc0d0f9279761e7e0a9e` - Hardware: Local Ubuntu / CUDA · Local Ubuntu / CPU - Row intervals: participant bootstrap, 10,000 draws - Paired-difference intervals: paired participant bootstrap, 10,000 draws One training run per cell with the matrix seed and the matrix recipe: AdamW, 20 epochs (sleep 10), no schedule, no early stopping. This is the recipe of the EEGNet cells; it is not tuned for any network and it is not the original papers' training. Intervals describe variation across people, not across training runs. ## Downloads - [Results · CSV ↓](https://bci.report/data/sleep-scalp-results.csv) - [Protocol · JSON ↓](https://bci.report/data/sleep-scalp-protocol.json) - [Core matrix · JSON ↓](https://bci.report/data/experiments.json) [This protocol's rows in core-matrix-update-20261106 · CSV ↓](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) · [Core matrix: rows run under its exact protocols · JSON ↓](https://bci.report/data/core-matrix-20261106.json) ### Other protocols - [Motor imagery & rest on ds003810](https://bci.report/protocols/mi-rest/) - [Idle & command on ds005342](https://bci.report/protocols/idle/) - [SSVEP · 8 channels on BETA](https://bci.report/protocols/beta-8ch/) - [SSVEP · 4 channels on BETA](https://bci.report/protocols/beta-4ch/) - [Arithmetic & rest on EEGMAT](https://bci.report/protocols/arithmetic-rest/) - [P300 target ERP on ds006593](https://bci.report/protocols/p300-target/) - [Semantic target ERP on TMNRED / ds005383](https://bci.report/protocols/semantic-target/) [All protocols →](https://bci.report/protocols/) · [The core matrix on the home page →](https://bci.report/#overview) · [All datasets →](https://bci.report/datasets/) · [All methods →](https://bci.report/methods/) ## Cite this page BCI Report (2026). *Sleep staging on EESM19 scalp subset*. https://bci.report/protocols/sleep-scalp/ Figures from releases [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106) (2026-11-06), [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106) (2026-11-06), [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106) (2026-11-06), [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013) (2026-10-13), [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004) (2026-10-04), [`research-preview-20260920`](https://bci.report/releases/#research-preview-20260920) (2026-09-20). Cite the upstream dataset as well: its credit is on this page. Every release is archived on Zenodo: [doi:10.5281/zenodo.23123296](https://doi.org/10.5281/zenodo.23123296). [BibTeX for the site and its releases →](https://bci.report/api/#cite-heading) · [CITATION.cff ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) --- Markdown copy of https://bci.report/protocols/sleep-scalp/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # BCI Report data API: static JSON and CSV at stable URLs Data API Every published result is a static file: the same bytes the pages are built from. ## How it works ### Stable URLs Every file lives at https://bci.report/data/. A file is only ever replaced by a reviewed release, and the release archive gives the SHA-256 of the bytes currently served. ### No key, no account Plain HTTPS GET. Static files behind a CDN; no rate limit is promised, so cache what you use. Any HTTP client can fetch the files under /data/, the llms files and the Markdown copies, so pd.read_csv can read a CSV straight from its URL. Files under /data/ allow cross-origin reads, so a notebook running in a browser can fetch them too. ### Licence The aggregate results are CC BY 4.0. The recordings they were computed from belong to their authors and keep their own terms — each dataset page names them. The code that builds the site and the exports is MIT-licensed. ### Mirror The same files are on Hugging Face as the dataset Twu31/bci-report, with the per-protocol tables merged into loadable configurations. ## Files | File | What it holds | Release | | --- | --- | --- | | [`open-register-20261106.json`](https://bci.report/data/open-register-20261106.json) | The open register, canonical: every EEG dataset we know of and every EEG model and paper we track, one row each under an opaque id, with its ladder (dataset or model), the furthest step checked, each step’s evidence as pass, gap or not applicable with its source, the gaps, its status, the terms as stated and by whom, aliases (repository ids, accessions, MOABB loaders), findings as codes with their parameters, and links; the step definitions in English and Chinese, the rules, the sources with the frozen query, the counts and the uploader families. | [`open-register-update-20261106`](https://bci.report/releases/archive/#open-register-update-20261106) | | [`open-register-20261106.csv`](https://bci.report/data/open-register-20261106.csv) | The open register’s dataset ladder, flat, for spreadsheets and the Hugging Face viewer: one row per dataset entry with its id, group, name, host, licence as stated, whether access is gated, the furthest step checked, gaps, status and page. Findings are in the JSON. | [`open-register-update-20261106`](https://bci.report/releases/archive/#open-register-update-20261106) | | [`core-matrix-20261106.json`](https://bci.report/data/core-matrix-20261106.json) | Rows of the core matrix run under its exact protocols (filter-bank CCA on both BETA protocols; TS+LR, MDM, xDAWN+LDA and XDAWNCov+TS+LR on their core protocols and idle; ShallowFBCSPNet, Deep4Net and CSP+LDA in cells the matrix left empty): full-precision values, recipes, roles, backends, the paired differences shown on the protocol and method pages, and every matrix cell marked released, in this file, not applicable or not run. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-mi-rest.csv`](https://bci.report/data/core-matrix-20261106-mi-rest.csv) | Motor imagery & rest: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-idle.csv`](https://bci.report/data/core-matrix-20261106-idle.csv) | Idle & command: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-beta-8ch.csv`](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) | SSVEP · 8 channels: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-beta-4ch.csv`](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) | SSVEP · 4 channels: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-arithmetic-rest.csv`](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) | Arithmetic & rest: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-p300-target.csv`](https://bci.report/data/core-matrix-20261106-p300-target.csv) | P300 target ERP: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-semantic-target.csv`](https://bci.report/data/core-matrix-20261106-semantic-target.csv) | Semantic target ERP: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`core-matrix-20261106-sleep-scalp.csv`](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) | Sleep staging: this file’s rows in the core results CSV’s columns and units, then release, recipe, role, backend and protocol line. | [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) | | [`logpower-ridge-update.json`](https://bci.report/data/logpower-ridge-update.json) | BCI Report’s log band-power ridge (5 bands, visible channels) beside the Spectral ridge row on seven core protocols: per protocol, balanced accuracy with its participant-bootstrap interval, macro F1, people, channels and chance level; its paired differences from the Spectral ridge row and from one named frozen-encoder row with their intervals and readings; the recipe and how it differs from the Spectral ridge row; two summary lines on its head and input scaling; rights; and what is not published. | [`logpower-ridge-update-20261106`](https://bci.report/releases/archive/#logpower-ridge-update-20261106) | | [`signal-jepa-eegsimpleconv-update.json`](https://bci.report/data/signal-jepa-eegsimpleconv-update.json) | Signal-JEPA (frozen encoder with a ridge head, its own 0.5–40 Hz, 128 Hz input) and EEGSimpleConv (trained from scratch, mean of three seeds) on motor imagery & rest and arithmetic & rest, beside the core matrix: balanced accuracy, macro F1, Signal-JEPA’s participant-bootstrap interval, each model’s difference from EEGNet on the same people, identities and licences by part, the directory correction, and what is not published. | [`signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/archive/#signal-jepa-eegsimpleconv-update-20261106) | | [`foundation-checkpoints-update.json`](https://bci.report/data/foundation-checkpoints-update.json) | LaBraM at Scale’s Quark-B, Quark-L, Quark-H, Atom-L, Molecule-L and Quark-G, with REVE Small, as frozen probes on the eight core protocols: each cell’s balanced accuracy with its participant-bootstrap interval, macro F1 and chance level (idle: trial counts and window metrics), the Quark ladder’s paired differences, pretraining exposure per cell, terms and limitations. | [`foundation-checkpoints-update-20261106`](https://bci.report/releases/archive/#foundation-checkpoints-update-20261106) | | [`random-init-controls-update.json`](https://bci.report/data/random-init-controls-update.json) | Random-initialization controls for REVE Large, ST-EEGFormer Base and Large, SingLEM and EEGMamba on motor imagery and rest, arithmetic and rest, and sleep staging: pretrained and random-mean balanced accuracy and their paired difference with its interval, under the fixed and the train-selected readout. | [`random-init-controls-update-20261106`](https://bci.report/releases/archive/#random-init-controls-update-20261106) | | [`nsp-components-update.json`](https://bci.report/data/nsp-components-update.json) | A local test of the components of Neural State Prediction (arXiv:2609.31167v1, cited only), in a panel of its own after the counted pilots on the pilot-studies page, in three blocks: a small model trained on EEGMAT with the residualised target on or off and the context geometry topological or random, with the contrasts fixed in advance, position-probe means and within-model content-dependence drops; the same model never trained against the trained cells at three fitting-label budgets, analysed after the main results, with a sampled replay; and the published eeg-fm-masking encoders and a log band-power ridge with channels zeroed on EEGMAT and ds003810, with the head kept or refitted. Equal-person means and paired differences with descriptive participant-bootstrap intervals; methods, audit scope, the first incomplete audit, what was not run, credits and what is not published. | [`nsp-components-update-20261106`](https://bci.report/releases/archive/#nsp-components-update-20261106) | | [`electrode-shift-update.json`](https://bci.report/data/electrode-shift-update.json) | What if the cap sits one electrode off? A simulation on BETA’s 70 people: six decoders fitted at the original eight posterior sites and tested with every electrode read one position over (forward one row, left or right one column), deployed and recalibrated; the twelve prespecified contrasts with Bonferroni intervals and audited counts of people losing or gaining 10 points or more; all 144 contrasts, the neighbour maps, the spline doses and the bridges to the published cells. | [`electrode-shift-update-20261106`](https://bci.report/releases/archive/#electrode-shift-update-20261106) | | [`eldbeta-cohort-update.json`](https://bci.report/data/eldbeta-cohort-update.json) | eldBETA checked end to end on all 100 published participants: training-free CCA and filter-bank CCA mean participant accuracy with paired participant-bootstrap intervals and chance, the paired difference, the audited counts at the one-person binomial threshold, the file checks and the located non-finite values, and the rights record. | [`eldbeta-cohort-update-20261106`](https://bci.report/releases/archive/#eldbeta-cohort-update-20261106) | | [`clinical-aseeg-update.json`](https://bci.report/data/clinical-aseeg-update.json) | ASEEG, schizophrenia and controls from resting EEG: age-and-sex comparators beside EEG spectra, EEG with age and sex, age-and-sex-residualised EEG and frozen CBraMod (balanced accuracy and AUROC with paired participant-bootstrap intervals, macro F1), in eyes-closed and eyes-open recordings of the same people; the primary and secondary contrasts; the claim boundary and the rights record. | [`clinical-aseeg-update-20261106`](https://bci.report/releases/archive/#clinical-aseeg-update-20261106) | | [`open-register-20261028.json`](https://bci.report/data/open-register-20261028.json) | The open register, canonical: every EEG dataset we know of and every EEG model and paper we track, one row each under an opaque id, with its ladder (dataset or model), the furthest step checked, each step’s evidence as pass, gap or not applicable with its source, the gaps, its status, the terms as stated and by whom, aliases (repository ids, accessions, MOABB loaders), findings as codes with their parameters, and links; the step definitions in English and Chinese, the rules, the sources with the frozen query, the counts and the uploader families. | [`open-register-update-20261028`](https://bci.report/releases/archive/#open-register-update-20261028) | | [`open-register-20261028.csv`](https://bci.report/data/open-register-20261028.csv) | The open register’s dataset ladder, flat, for spreadsheets and the Hugging Face viewer: one row per dataset entry with its id, group, name, host, licence as stated, whether access is gated, the furthest step checked, gaps, status and page. Findings are in the JSON. | [`open-register-update-20261028`](https://bci.report/releases/archive/#open-register-update-20261028) | | [`ssvep-followups-update.json`](https://bci.report/data/ssvep-followups-update.json) | More audited follow-ups on SSVEP frequencies a decoder never trained on, cards 8 to 14 of that question page, on route 3’s BETA cohort and, for some, on Wearable-102’s dry and wet headbands: the endpoints with their chance levels; a name key from the site’s names to the export codes, and the estimates several cards share; for each study its design, levels and paired contrasts with their intervals and interval method, roles and flags, its checks as pass or fail with counts, and what was not run; which intervals are adjusted; CPU timings of each component (warm, single-thread, never summed); rights records, the filter-bank CCA credit and the audit record. | [`ssvep-followups-update-20261028`](https://bci.report/releases/archive/#ssvep-followups-update-20261028) | | [`neurorvq-update.json`](https://bci.report/data/neurorvq-update.json) | NeuroRVQ’s frozen backbone on three existing tasks (EEGMAT’s arithmetic and rest, motor imagery against rest, P300 target against non-target), a separate pilot panel: for each task the cohort, split and input chain, the participant-mean balanced accuracy and macro-F1 of the backbone and of simple features of the same windows under one fixed ridge head, and one paired difference with its participant-bootstrap interval; the checkpoint’s identity and how it was restored, pretraining exposure (unknown) with its sources, the protocols not run and why, the terms of the code, the weights and each dataset, limitations, the audit record and what is not published. | [`neurorvq-update-20261028`](https://bci.report/releases/archive/#neurorvq-update-20261028) | | [`eegmat-ten-second-update.json`](https://bci.report/data/eegmat-ten-second-update.json) | Two frozen encoders, EEG-Captioner’s and AXON, on ten-second windows of EEGMAT’s arithmetic and rest, a separate pilot panel with a subsection for each: the cohort and its preprocessing, the source annotation-QC totals printed beside every score, each encoder’s participant-mean balanced accuracy against Welch band power on the same windows with a paired participant-bootstrap interval, each encoder’s identity and terms, AXON’s own numerical input check, why no difference between the two encoders is computed, limitations, the audit record and what is not published. | [`eegmat-ten-second-update-20261028`](https://bci.report/releases/archive/#eegmat-ten-second-update-20261028) | | [`dataset-checks-update.json`](https://bci.report/data/dataset-checks-update.json) | Four public EEG datasets checked end to end from their files. Kim beta-range SSVEP, a worked case: training-free CCA on all its participants, the mean participant accuracy with its participant-bootstrap interval against chance, the intake and how it was checked, caveats, and a hold on amplitude-dependent models. MIND, a conditional case on record IDs: CSP + LDA beside an ordinal lookup and a majority control for two stored-code pairs, record-ID resampling ranges, the label schedules, a separate secondary panel and the coverage chain. Damm and NeBULA, one-person status cases: design, counts, file facts and the arms run, with every score held. For each source its rights record, its licence statements as each source states them, its consent and ethics statements, and the holds. | [`dataset-checks-update-20261028`](https://bci.report/releases/archive/#dataset-checks-update-20261028) | | [`open-register-20261020.json`](https://bci.report/data/open-register-20261020.json) | The open register, canonical: every EEG dataset we know of and every EEG model and paper we track, one row each under an opaque id, with its ladder (dataset or model), the furthest step checked, each step’s evidence as pass, gap or not applicable with its source, the gaps, its status, the terms as stated and by whom, aliases (repository ids, accessions, MOABB loaders), findings as codes with their parameters, and links; the step definitions in English and Chinese, the rules, the sources with the frozen query, the counts and the uploader families. | [`open-register-update-20261020`](https://bci.report/releases/archive/#open-register-update-20261020) | | [`open-register-20261020.csv`](https://bci.report/data/open-register-20261020.csv) | The open register’s dataset ladder, flat, for spreadsheets and the Hugging Face viewer: one row per dataset entry with its id, group, name, host, licence as stated, whether access is gated, the furthest step checked, gaps, status and page. Findings are in the JSON. | [`open-register-update-20261020`](https://bci.report/releases/archive/#open-register-update-20261020) | | [`new-ssvep-frequencies-update.json`](https://bci.report/data/new-ssvep-frequencies-update.json) | What lets a decoder answer an SSVEP frequency it never trained on? Seven audited follow-ups to route 3 on its BETA cohort, and the same models run unchanged on Wearable-102: the endpoints with their chance levels; a name key from the site’s names to the export codes; for each study its levels and paired contrasts with their intervals and interval method, roles and flags, its checks as pass or fail with counts, and what was not run; the stimulus phases taken from the BETA authors’ corrigendum; a status card for the cross-spectral round, with no figures; rights records and the audit record. | [`new-ssvep-frequencies-update-20261020`](https://bci.report/releases/archive/#new-ssvep-frequencies-update-20261020) | | [`data-quality-update.json`](https://bci.report/data/data-quality-update.json) | Three public EEG datasets checked end to end from their files. Shin2022, the worked case: a participant-held-out diagnostic with three arms (a training prior, a normalised spectral ridge and a retrospective duration control; participant-mean balanced accuracy with its participant-bootstrap interval, and a pooled secondary value), the coverage chain from the runs in the release to the signal-valid windows, where the catalog’s trial length comes from, how the files were checked, its holds, rights and consent and ethics statements. eldBETA: counts for one participant’s recordings, with the score held. ZuCo 2.0: a timing hold on one converted run, with no figure. | [`data-quality-update-20261020`](https://bci.report/releases/archive/#data-quality-update-20261020) | | [`clinical-iesseeg-update.json`](https://bci.report/data/clinical-iesseeg-update.json) | IESSEEG, an infant case-control pilot with age and recording-setting controls: six arms with and without EEG (a training prior, age only, recording settings only, EEG spectral features, EEG with age, and EEG after training-fold age residualization) and four arms that read frozen NeuroLM features with the same probes. Out-of-fold point estimates of balanced accuracy, AUROC and secondary metrics with confusion counts, for everyone and for the people aged 2 or under; the cohort and its input check; the release card’s attribution and ethics statement; the audit boundary, limitations and what is not published; and the clinical-groups page’s medical disclaimer. | [`clinical-iesseeg-update-20261020`](https://bci.report/releases/archive/#clinical-iesseeg-update-20261020) | | [`neurolm-eegmat-update.json`](https://bci.report/data/neurolm-eegmat-update.json) | NeuroLM’s frozen features on EEGMAT’s arithmetic and rest, a separate pilot panel: equal-person balanced accuracy and macro-F1 of three representations (tokenizer features, GPT hidden states and band power) under one fixed ridge head, three paired differences with their participant-bootstrap intervals, the adapter and the probe, the checkpoint’s identity and the terms of each part, pretraining exposure (unknown) with its sources, limitations, the audit record and what is not published. | [`neurolm-eegmat-update-20261020`](https://bci.report/releases/archive/#neurolm-eegmat-update-20261020) | | [`eegpt-core-update.json`](https://bci.report/data/eegpt-core-update.json) | EEGPT beside the core matrix, from the same runs as the published LaBraM rows and, except on idle, the CBraMod rows: its eight core-matrix cells as a frozen row (balanced accuracy with its participant-bootstrap interval, macro F1, chance level and pretraining exposure per cell, each percent figure also in the core rows’ units; idle false activations, detections within 3 s and people who always abstained as counts, with the window balanced accuracy and AUROC), the footnotes beside the cells they bear on, how each cell stands against the published rows (above, below or overlapping by the 95% intervals, with the gap in percentage points), its pretraining-attribution rows and paired contrasts on mi-rest and EEGMAT in the published deployment-topics fields, the model’s terms (checkpoint licence unresolved, no weights distributed, Apache-2.0 the code’s alone), the owner’s release record and what is not published. | [`eegpt-core-update-20261013`](https://bci.report/releases/archive/#eegpt-core-update-20261013) | | [`pilot-studies-update.json`](https://bci.report/data/pilot-studies-update.json) | Small pilots outside every leaderboard and model ranking. The Orthosis error-detection dataset as one cohort in three results: a 64-channel classical EEG baseline (balanced accuracy, AUROC and average precision with person-bootstrap intervals, a direction/time confound probe, the training majority and class prior, the matched EEG-minus-confound difference with how many people were above, permutation-control means with observed ranges, and the response-clean subset as counts), the same baseline on 58 channels, and frozen EEGPT against those matched classical features (the paired difference with its interval, the checkpoint’s identity and licence status; no weights). BINGO Greek and English inner speech within and across languages (balanced and top-3 accuracy against their priors, order-only, high-frequency and label-shuffle controls, the cross-minus-within contrast, the protocol and the project’s documented preprocessing). A mask-aware EEG analysis of a handwriting competition dataset, five people (balanced accuracy of relative EEG and three controls that see no EEG, with person-bootstrap intervals except the mask-only control’s, three paired contrasts, label-shuffle means with observed ranges, channel availability over trials and its flag counts, and the earlier analysis of the same people as context). REFED’s stored-clip diagnostic (equal-axis mean absolute error for EEG and three source-only controls in two conditions). Status-only entries, and each source’s rights, consent and ethics record, with the handwriting dataset’s stated gap. | [`pilot-studies-update-20261012`](https://bci.report/releases/archive/#pilot-studies-update-20261012) | | [`beyond-scalp-update.json`](https://bci.report/data/beyond-scalp-update.json) | Do simple baselines work on signals other than scalp EEG? Human fNIRS on Bak et al.’s finger and foot tapping, a person left out at a time (participant-macro balanced accuracy with its descriptive interval, class recalls, and a training-prior comparator); one macaque’s motor-cortex spiking on MC_RTT (a ridge decoder of finger velocity against the training mean: R², RMSE and MAE, including the metric on which the ridge is worse, with trial, neuron and row counts); a numerical coverage check of ds007420’s fNIRS recordings (counts only, with the gates that hold its classifier); and two sources prepared but not scored, FALCON H2 and Rabut et al.’s functional ultrasound, as counts. Species and modality on every entry; each source’s rights record, with its consent and ethics statement or, for the macaque, its animal-care statement. | [`beyond-scalp-update-20261012`](https://bci.report/releases/archive/#beyond-scalp-update-20261012) | | [`questions-in-language-update.json`](https://bci.report/data/questions-in-language-update.json) | Route 3, questions in language, on BETA SSVEP and BOAS sleep: one question-conditioned head asked each question by a question number, a label template or a description — the 35 pre-declared entries with n, seeds, interval, both flags, the route-sentence rule and the route sentence, kept apart as seen questions, rewordings and questions the EEG head was never trained on; per-arm levels with chance-corrected values, generalisation costs, read-offs and flip rates; the pre-run checks as pass or fail and counts, with the two failed canary revisions and the owner’s override disclosed; the secondary results (EESM19, OpenBMI, REVE-L, a second text encoder, Chinese wordings, negation, extrapolation, Wearable-102), labelled single-seed and likely inconclusive where declared; BOAS’s three stated gaps and attribution, rights records, references and the audit record. | [`questions-in-language-update-20261008`](https://bci.report/releases/archive/#questions-in-language-update-20261008) | | [`questions-in-language-wordings.json`](https://bci.report/data/questions-in-language-wordings.json) | The wordings route 3 asked its questions with, published with its results: 787 texts in English and Chinese for sleep, SSVEP and motor imagery (training and held-out sentence frames, label names and synonyms, training and held-out descriptions, descriptions and names of the unseen questions, negation frames), written before any score by two agents from a fixed brief; the leakage rules they passed; the held-out frequency partition of the five rotations and the extrapolation band; the declared derangements of the shuffled control; the numeric frequency code; and the pinned text encoders. Text only, CC BY 4.0. | [`questions-in-language-update-20261008`](https://bci.report/releases/archive/#questions-in-language-update-20261008) | | [`later-sessions-update.json`](https://bci.report/data/later-sessions-update.json) | Does a decoder trained on an earlier session still work later? WBCIC-SHU session 1 to session 3 in two separate cohorts (a source-majority prior and a relative spectral ridge, and frozen CBraMod with a session-1 ridge readout: balanced accuracy with its participant-bootstrap interval, accuracy, macro F1 and the paired differences); the longitudinal RSVP source’s first-visit ERP baseline at the nominal Day 7, 80 and 200 visits (AUROC with its interval, average precision, log loss, Brier score and ECE, event counts, the paired Day 200 minus Day 7 contrast and how many people declined); and Forenzo’s continuous cursor tracking, a negative result (per cohort and response arm, the ridge’s mean and median error against the source-mean comparator, paired differences, how many records the ridge did worse on, secondary metrics with nulls kept null, coverage and holds); with each source’s rights, consent and ethics record and its limitations. | [`later-sessions-update-20261008`](https://bci.report/releases/archive/#later-sessions-update-20261008) | | [`shared-representation-update.json`](https://bci.report/data/shared-representation-update.json) | Route 2, one representation and several questions, on OpenBMI motor imagery and BOAS sleep: independent models, fixed linear heads on a shared encoder, a shared hidden layer and the question-conditioned head, at matched data and compute, with EEGNet from scratch, frozen CBraMod features and CBraMod adapted by LoRA — every paired contrast with its interval, its difference and margin flags and its gate; per-arm balanced accuracy, AUROC and log loss; the route sentence per domain and the cost ledger; the derivable-question test (log R); a crude EESM19 replication; BOAS’s three stated gaps and attribution, rights records and the audit record. | [`shared-representation-update-20261007`](https://bci.report/releases/archive/#shared-representation-update-20261007) | | [`foundation-models-update.json`](https://bci.report/data/foundation-models-update.json) | Further foundation encoders beside the core matrix: sixteen encoder checkpoints of eleven further models as frozen probes on the eight core protocols (balanced accuracy with its participant-bootstrap interval, macro F1, chance flag and pretraining exposure per cell; idle false activations and detections as counts; two cells not run, with the reason), the masking ablation, nine models adapted on EEGMAT (head on a frozen encoder against rank-4 LoRA: three-seed means, per-seed means, paired changes with people helped and harmed, trainable parameters), comparisons decided by interval overlap, the pretraining-exposure table with sources, weights licences, row footnotes and the audit record. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-mi-rest.csv`](https://bci.report/data/foundation-models-mi-rest.csv) | Motor imagery & rest: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-idle.csv`](https://bci.report/data/foundation-models-idle.csv) | Idle & command: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-beta-8ch.csv`](https://bci.report/data/foundation-models-beta-8ch.csv) | SSVEP · 8 channels: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-beta-4ch.csv`](https://bci.report/data/foundation-models-beta-4ch.csv) | SSVEP · 4 channels: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-arithmetic-rest.csv`](https://bci.report/data/foundation-models-arithmetic-rest.csv) | Arithmetic & rest: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-p300-target.csv`](https://bci.report/data/foundation-models-p300-target.csv) | P300 target ERP: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-semantic-target.csv`](https://bci.report/data/foundation-models-semantic-target.csv) | Semantic target ERP: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`foundation-models-sleep-scalp.csv`](https://bci.report/data/foundation-models-sleep-scalp.csv) | Sleep staging: the further foundation-encoder rows — one per encoder checkpoint, in the core results CSV’s columns and units, then status, panel, chance flag, pretraining exposure, licence and footnote. Not run is empty with its reason. | [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) | | [`reliable-decisions-update.json`](https://bci.report/data/reliable-decisions-update.json) | Route 1, reliable decisions, on EEGMAT and BETA: each method’s coverage and error under a fixed confidence threshold (with how many people had nothing accepted), a coverage target and a certified selective risk (with folds certified and folds over target); the learned reject option against calibrated confidence (AURC, error at 80% coverage); errors at a matched coverage; NLL, ECE and Brier score; and person-specific recalibration with its label cost — with paired person-bootstrap intervals and verdicts. A crude ds003810 panel, secondary seeds, sensitivity arms and the audit record. | [`reliable-decisions-update-20261004`](https://bci.report/releases/archive/#reliable-decisions-update-20261004) | | [`large-source-update.json`](https://bci.report/data/large-source-update.json) | Dreem sleep staging in two separate cohorts, DOD-H (healthy) and DOD-O (obstructive sleep apnoea): a training prior and a spectral ridge with accuracy, balanced accuracy, macro F1 and Cohen’s kappa, the paired balanced-accuracy gain and the ridge’s per-stage recall, precision and F1 (nulls kept null); and OpenBMI cross-session motor-imagery calibration, 51 people at 0, 10, 20 and 40 labelled session-2 trials, with paired changes and how many people declined. Rights records and one figure-free hold. | [`large-source-update-20261003`](https://bci.report/releases/archive/#large-source-update-20261003) | | [`extension-update.json`](https://bci.report/data/extension-update.json) | Twenty further people of the asynchronous SSVEP release under a pilot-fixed and a personal rejection threshold (detection, coverage, correct-and-accepted, false acceptance per state, people helped and harmed), and LTRSVP image-rate and recording transfer (two primary arms, paired difference, 3×3 matrix), with rights records. | [`extension-update-20261002`](https://bci.report/releases/archive/#extension-update-20261002) | | [`adaptation-update.json`](https://bci.report/data/adaptation-update.json) | LaBraM on EEGMAT, new people, same task, zero labels from the test person, adapted three ways (head only, last block, rank-4 LoRA) over three seeds: scores, paired changes with people helped or harmed, trainable parameters and training time; the next-day experiment as status only. | [`adaptation-update-20261001`](https://bci.report/releases/archive/#adaptation-update-20261001) | | [`context-update.json`](https://bci.report/data/context-update.json) | Screen-to-VR P300 transfer, treadmill walking speed beside a movement-nuisance comparator, and the asynchronous SSVEP non-control pilot, with audits. | [`context-update-20260927`](https://bci.report/releases/archive/#context-update-20260927) | | [`clinical-update.json`](https://bci.report/data/clinical-update.json) | Parkinson's disease vs. controls from resting-state EEG (ds004584) beside an age-and-sex-only comparator, the claim boundary, and status-only sources. | [`clinical-update-20260923`](https://bci.report/releases/archive/#clinical-update-20260923) | | [`evidence-update.json`](https://bci.report/data/evidence-update.json) | In-ear vs. scalp sleep staging (EESM23), four vs. sixteen electrodes (Alpha Waves), and the physical head phantom, each with its rights record. | [`evidence-update-20260922`](https://bci.report/releases/archive/#evidence-update-20260922) | | [`deployment-topics.json`](https://bci.report/data/deployment-topics.json) | Deployment topics: dry vs. wet sensor transfer, calibration budget, movement (SSVEP and ERP) and pretraining controls — measurements, paired contrasts, seed sensitivity and dataset citations. | [`deployment-topics-20260920-v1`](https://bci.report/releases/archive/#deployment-topics-20260920-v1) | | [`experiments.json`](https://bci.report/data/experiments.json) | The core matrix: every protocol with each method’s score, interval, cohort, electrode count, training mode and limitations, plus the model and dataset directories. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`mi-rest-results.csv`](https://bci.report/data/mi-rest-results.csv) | Motor imagery & rest: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`mi-rest-protocol.json`](https://bci.report/data/mi-rest-protocol.json) | Motor imagery & rest: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`idle-results.csv`](https://bci.report/data/idle-results.csv) | Idle & command: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`idle-protocol.json`](https://bci.report/data/idle-protocol.json) | Idle & command: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`beta-8ch-results.csv`](https://bci.report/data/beta-8ch-results.csv) | SSVEP · 8 channels: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`beta-8ch-protocol.json`](https://bci.report/data/beta-8ch-protocol.json) | SSVEP · 8 channels: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`beta-4ch-results.csv`](https://bci.report/data/beta-4ch-results.csv) | SSVEP · 4 channels: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`beta-4ch-protocol.json`](https://bci.report/data/beta-4ch-protocol.json) | SSVEP · 4 channels: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`arithmetic-rest-results.csv`](https://bci.report/data/arithmetic-rest-results.csv) | Arithmetic & rest: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`arithmetic-rest-protocol.json`](https://bci.report/data/arithmetic-rest-protocol.json) | Arithmetic & rest: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`p300-target-results.csv`](https://bci.report/data/p300-target-results.csv) | P300 target ERP: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`p300-target-protocol.json`](https://bci.report/data/p300-target-protocol.json) | P300 target ERP: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`semantic-target-results.csv`](https://bci.report/data/semantic-target-results.csv) | Semantic target ERP: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`semantic-target-protocol.json`](https://bci.report/data/semantic-target-protocol.json) | Semantic target ERP: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`sleep-scalp-results.csv`](https://bci.report/data/sleep-scalp-results.csv) | Sleep staging: one row per method — score, interval, secondary metric, cohort and compute time. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | | [`sleep-scalp-protocol.json`](https://bci.report/data/sleep-scalp-protocol.json) | Sleep staging: the protocol — cohort, channels, windows, split, training budget, rights and limitations. | [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) | [Release archive, with the SHA-256 of every file →](https://bci.report/releases/archive/) ## Examples ### Shell ``` curl -sO https://bci.report/data/experiments.json curl -s https://bci.report/data/context-update.json | jq '.results | keys' ``` ### Python ``` import pandas as pd, requests matrix = requests.get("https://bci.report/data/experiments.json", timeout=30).json() for track in matrix["tracks"]: print(track["id"], track["chanceLevel"], len(track["rows"])) mi_rest = pd.read_csv("https://bci.report/data/mi-rest-results.csv") ``` ### Hugging Face datasets ``` from datasets import load_dataset results = load_dataset("Twu31/bci-report", "results") # also: topics, contrasts, seed_sensitivity, foundation_models, eegpt_core, core_matrix_rows, core_matrix_contrasts, register ``` ## Cite Cite the site and the release you used (the current one is open-register-update-20261106), and the upstream dataset each figure was computed on. ``` @misc{bcireport, title = {BCI Report: public EEG decoding results reported with their protocol}, author = {{BCI Report}}, year = {2026}, howpublished = {\url{https://bci.report}}, doi = {10.5281/zenodo.23123296}, note = {Release open-register-update-20261106} } ``` [Upstream datasets and their credits →](https://bci.report/datasets/) [CITATION.cff, the same citation machine-readable ↗](https://raw.githubusercontent.com/twu3202/bci-report/main/CITATION.cff) · [Code on GitHub (MIT) ↗](https://github.com/twu3202/bci-report) · [Mirror on Hugging Face ↗](https://huggingface.co/datasets/Twu31/bci-report) ## For agents and crawlers - [`/llms.txt`](https://bci.report/llms.txt) — an index of every page with its short answer - [`/llms-full.txt`](https://bci.report/llms-full.txt) — every English page as Markdown, in one file - Every page has a Markdown copy: add index.md to its URL (e.g. /topics/dry-vs-wet/index.md) - [`/sitemap.xml`](https://bci.report/sitemap.xml) — every page, both languages, with last-modified dates - [`/releases.xml`](https://bci.report/releases.xml) — an Atom feed with one entry per reviewed release [BCI Report Explorer on Hugging Face ↗](https://huggingface.co/spaces/Twu31/bci-report-explorer): browse the same results, or query them through MCP tools. It reads one pinned release of the mirror and names it in every answer. --- Markdown copy of https://bci.report/api/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Releases Release log What the newest releases changed, with links to the pages and files. Every release’s full record, with its review manifest and the SHA-256 of each file it ships, is in [the archive](https://bci.report/releases/archive/). ## Newest releases ### 6 November 2026 - [Filter-bank CCA, Riemannian and xDAWN baselines, ShallowFBCSPNet, Deep4Net and CSP+LDA](https://bci.report/protocols/) join the [core matrix](https://bci.report/#overview) under its exact protocols; cells that do not apply are marked apart from cells not run. - Beside the matrix: [LaBraM at Scale and REVE Small](https://bci.report/topics/does-pretraining-help/#more-checkpoints) as frozen probes on every core protocol, random-initialization controls for five checkpoints, [Signal-JEPA and EEGSimpleConv](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) on two, and a log band-power ridge on seven. - A simulated [cap shift](https://bci.report/topics/fewer-electrodes/#cap-shift) on BETA; a local test of NSP’s target components in [a panel of its own after the counted pilots](https://bci.report/topics/pilot-studies/#nsp-components); eldBETA’s 100 participants scored without training; ASEEG beside age and sex. - [`open-register-update-20261106`](https://bci.report/releases/archive/#open-register-update-20261106) · [open-register-20261106.json](https://bci.report/data/open-register-20261106.json) · [open-register-20261106.csv](https://bci.report/data/open-register-20261106.csv) - [`core-matrix-update-20261106`](https://bci.report/releases/archive/#core-matrix-update-20261106) · [core-matrix-20261106.json](https://bci.report/data/core-matrix-20261106.json) · [core-matrix-20261106-mi-rest.csv](https://bci.report/data/core-matrix-20261106-mi-rest.csv) · [core-matrix-20261106-idle.csv](https://bci.report/data/core-matrix-20261106-idle.csv) · [core-matrix-20261106-beta-8ch.csv](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) · [core-matrix-20261106-beta-4ch.csv](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) · [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) · [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) · [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) · [core-matrix-20261106-sleep-scalp.csv](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) - [`logpower-ridge-update-20261106`](https://bci.report/releases/archive/#logpower-ridge-update-20261106) · [logpower-ridge-update.json](https://bci.report/data/logpower-ridge-update.json) - [`signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/archive/#signal-jepa-eegsimpleconv-update-20261106) · [signal-jepa-eegsimpleconv-update.json](https://bci.report/data/signal-jepa-eegsimpleconv-update.json) - [`foundation-checkpoints-update-20261106`](https://bci.report/releases/archive/#foundation-checkpoints-update-20261106) · [foundation-checkpoints-update.json](https://bci.report/data/foundation-checkpoints-update.json) - [`random-init-controls-update-20261106`](https://bci.report/releases/archive/#random-init-controls-update-20261106) · [random-init-controls-update.json](https://bci.report/data/random-init-controls-update.json) - [`nsp-components-update-20261106`](https://bci.report/releases/archive/#nsp-components-update-20261106) · [nsp-components-update.json](https://bci.report/data/nsp-components-update.json) - [`electrode-shift-update-20261106`](https://bci.report/releases/archive/#electrode-shift-update-20261106) · [electrode-shift-update.json](https://bci.report/data/electrode-shift-update.json) - [`eldbeta-cohort-update-20261106`](https://bci.report/releases/archive/#eldbeta-cohort-update-20261106) · [eldbeta-cohort-update.json](https://bci.report/data/eldbeta-cohort-update.json) - [`clinical-aseeg-update-20261106`](https://bci.report/releases/archive/#clinical-aseeg-update-20261106) · [clinical-aseeg-update.json](https://bci.report/data/clinical-aseeg-update.json) ### 28 October 2026 - Seven more follow-ups on [SSVEP frequencies a decoder never trained on](https://bci.report/topics/new-ssvep-frequencies/#cross-spectral), among them filter-bank CCA, shorter windows, fewer electrodes and fixed harmonic weights, each beside its untrained control. - Frozen-encoder probes in panels of their own: [NeuroRVQ](https://bci.report/topics/pilot-studies/#neurorvq) on arithmetic, motor imagery and P300, and [EEG-Captioner and AXON](https://bci.report/topics/pilot-studies/#eegmat-ten-second) on ten-second arithmetic and rest windows. - Four more datasets [checked end to end](https://bci.report/topics/data-quality/), with a CCA baseline on Kim’s beta-range SSVEP, and [the open register](https://bci.report/methods/register/) extended to match. - [`open-register-update-20261028`](https://bci.report/releases/archive/#open-register-update-20261028) · [open-register-20261028.json](https://bci.report/data/open-register-20261028.json) · [open-register-20261028.csv](https://bci.report/data/open-register-20261028.csv) - [`ssvep-followups-update-20261028`](https://bci.report/releases/archive/#ssvep-followups-update-20261028) · [ssvep-followups-update.json](https://bci.report/data/ssvep-followups-update.json) - [`neurorvq-update-20261028`](https://bci.report/releases/archive/#neurorvq-update-20261028) · [neurorvq-update.json](https://bci.report/data/neurorvq-update.json) - [`eegmat-ten-second-update-20261028`](https://bci.report/releases/archive/#eegmat-ten-second-update-20261028) · [eegmat-ten-second-update.json](https://bci.report/data/eegmat-ten-second-update.json) - [`dataset-checks-update-20261028`](https://bci.report/releases/archive/#dataset-checks-update-20261028) · [dataset-checks-update.json](https://bci.report/data/dataset-checks-update.json) ### 20 October 2026 - The open register: [every EEG dataset we know of](https://bci.report/datasets/register/) and [every EEG model we track](https://bci.report/methods/register/), in 527 rows, each at the furthest step checked, with its gaps listed. - Two new questions: what lets a decoder answer [SSVEP frequencies it never trained on](https://bci.report/topics/new-ssvep-frequencies/), and [what public EEG files show when checked end to end](https://bci.report/topics/data-quality/). - [IESSEEG](https://bci.report/topics/clinical-groups/#iesseeg) infant case-control results with age and recording-setting controls, and [NeuroLM’s frozen features on EEGMAT](https://bci.report/topics/pilot-studies/#neurolm-eegmat) in a panel of their own. - [`open-register-update-20261020`](https://bci.report/releases/archive/#open-register-update-20261020) · [open-register-20261020.json](https://bci.report/data/open-register-20261020.json) · [open-register-20261020.csv](https://bci.report/data/open-register-20261020.csv) - [`new-ssvep-frequencies-update-20261020`](https://bci.report/releases/archive/#new-ssvep-frequencies-update-20261020) · [new-ssvep-frequencies-update.json](https://bci.report/data/new-ssvep-frequencies-update.json) - [`data-quality-update-20261020`](https://bci.report/releases/archive/#data-quality-update-20261020) · [data-quality-update.json](https://bci.report/data/data-quality-update.json) - [`clinical-iesseeg-update-20261020`](https://bci.report/releases/archive/#clinical-iesseeg-update-20261020) · [clinical-iesseeg-update.json](https://bci.report/data/clinical-iesseeg-update.json) - [`neurolm-eegmat-update-20261020`](https://bci.report/releases/archive/#neurolm-eegmat-update-20261020) · [neurolm-eegmat-update.json](https://bci.report/data/neurolm-eegmat-update.json) ## Earlier releases - 13 October 2026 · [`eegpt-core-update-20261013`](https://bci.report/releases/archive/#eegpt-core-update-20261013) — [Core matrix (home page)](https://bci.report/) · [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) - 12 October 2026 · [`pilot-studies-update-20261012`](https://bci.report/releases/archive/#pilot-studies-update-20261012) — [Pilot studies](https://bci.report/topics/pilot-studies/) · [Methods](https://bci.report/methods/) - 12 October 2026 · [`beyond-scalp-update-20261012`](https://bci.report/releases/archive/#beyond-scalp-update-20261012) — [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/) - 8 October 2026 · [`questions-in-language-update-20261008`](https://bci.report/releases/archive/#questions-in-language-update-20261008) — [Questions in language](https://bci.report/topics/questions-in-language/) · [When not to act](https://bci.report/topics/when-not-to-act/) - 8 October 2026 · [`later-sessions-update-20261008`](https://bci.report/releases/archive/#later-sessions-update-20261008) — [Later sessions](https://bci.report/topics/later-sessions/) - 7 October 2026 · [`shared-representation-update-20261007`](https://bci.report/releases/archive/#shared-representation-update-20261007) — [One model, several questions](https://bci.report/topics/shared-encoder/) · [When not to act](https://bci.report/topics/when-not-to-act/) - 4 October 2026 · [`foundation-models-update-20261004`](https://bci.report/releases/archive/#foundation-models-update-20261004) — [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) - 4 October 2026 · [`reliable-decisions-update-20261004`](https://bci.report/releases/archive/#reliable-decisions-update-20261004) — [When not to act](https://bci.report/topics/when-not-to-act/) - 3 October 2026 · [`large-source-update-20261003`](https://bci.report/releases/archive/#large-source-update-20261003) — [Sleep-stage balance](https://bci.report/topics/sleep-staging/) · [How much calibration?](https://bci.report/topics/calibration-budget/) - 2 October 2026 · [`extension-update-20261002`](https://bci.report/releases/archive/#extension-update-20261002) — [When not to act](https://bci.report/topics/when-not-to-act/) · [Screen to VR](https://bci.report/topics/screen-to-vr/) - 1 October 2026 · [`adaptation-update-20261001`](https://bci.report/releases/archive/#adaptation-update-20261001) — [Which part to update?](https://bci.report/topics/model-adaptation/) - 27 September 2026 · [`context-update-20260927`](https://bci.report/releases/archive/#context-update-20260927) — [Screen to VR](https://bci.report/topics/screen-to-vr/) · [On the move](https://bci.report/topics/on-the-move/) · [When not to act](https://bci.report/topics/when-not-to-act/) - 23 September 2026 · [`clinical-update-20260923`](https://bci.report/releases/archive/#clinical-update-20260923) — [Clinical groups](https://bci.report/topics/clinical-groups/) - 22 September 2026 · [`evidence-update-20260922`](https://bci.report/releases/archive/#evidence-update-20260922) — [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) · [On the move](https://bci.report/topics/on-the-move/) · [Which part to update?](https://bci.report/topics/model-adaptation/) - 20 September 2026 · [`deployment-topics-20260920-v1`](https://bci.report/releases/archive/#deployment-topics-20260920-v1) — [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) · [On the move](https://bci.report/topics/on-the-move/) · [How much calibration?](https://bci.report/topics/calibration-budget/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) - 20 September 2026 · [`research-preview-20260920`](https://bci.report/releases/archive/#research-preview-20260920) — [Core matrix (home page)](https://bci.report/) ## Corrections Each correction to released wording is also made on the page that shows it; the released files keep their original bytes, whose SHA-256 values are in [the release archive](https://bci.report/releases/archive/). | Date | Files | Correction | | --- | --- | --- | | 6 November 2026 | `experiments.json` · `arithmetic-rest-protocol.json` | The arithmetic-rest protocol’s limitations do not say that EEGMAT records rest before the task for every participant, as PhysioNet describes it. The protocol page states it beside the other limitations. [See the page →](https://bci.report/protocols/arithmetic-rest/#recording-order) | | 6 November 2026 | `experiments.json` | The model directory’s notes in experiments.json say Signal-JEPA was not downloaded or executed, and EEGSimpleConv not executed, for that release. Both were run the same day in a separate study on motor imagery & rest and arithmetic & rest; their results are on those two protocol pages. [See the page →](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) | | 1 October 2026 | `deployment-topics.json` | The ensemble-TRCA rows call the original method a three-filter-bank experiment. The original paper and the reference code’s tutorials use five sub-bands; three is only the default of the code’s functions. Our runs used one fixed band either way, and no score changes. [See the page →](https://bci.report/topics/calibration-budget/#methods-and-limits) | | 1 October 2026 | `experiments.json` · `mi-rest-results.csv` · `mi-rest-protocol.json` · `deployment-topics.json` | The ds003810 credit links only the 2022 Data in Brief description. The OpenNeuro record asks users to cite Peterson, Galván, Hernández and Spies, Heliyon 6(3):e03425 (2020); the dataset page now gives both. [See the page →](https://bci.report/datasets/ds003810/#methods-and-limits) | ## Holds register Each item held back, with the date its hold was opened and, where it was, resolved. | Item | Opened | Status | Why | | --- | --- | --- | --- | | [Kim beta-range SSVEP · amplitude-dependent models](https://bci.report/topics/data-quality/#kim2025-beta-range) | 28 October 2026 | Open | Amplitude units, channel gains and the saved reference history are not established, so models whose output depends on amplitude are held on these files. Training-free CCA, which does not depend on a common amplitude scale, is published for all 40 participants. | | [Damm finger motor imagery · one-person score](https://bci.report/topics/data-quality/#damm) | 28 October 2026 | Open | One participant’s two random-order runs, checked end to end: 150 delivered cues, and eight variants and controls fixed before scoring and independently audited. Every score is held under the one-person rule, by the owner’s decision; design, counts and file facts are published. | | [NeBULA · one-person scores](https://bci.report/topics/data-quality/#nebula) | 28 October 2026 | Open | One participant reaching to three targets under free, low and high assistance: 89 delivered trials, with EEG, EMG and EEG plus EMG arms run and independently audited. Every score is held under the one-person rule, by the owner’s decision; design, counts and file facts are published. | | [ZuCo on NEMAR · converted-run timing](https://bci.report/topics/data-quality/#zuco2) | 20 October 2026 | Open | In one converted run of ZuCo 2.0 on NEMAR, some fixation times fall outside the recording. No EEG–text score is computed until the converter’s timing rules and a source-checksummed fixture are available; no timing was corrected and no row removed. | | [ds007420 · fNIRS classifier](https://bci.report/topics/beyond-scalp-eeg/#ds007420) | 12 October 2026 | Open | Its numeric task codes have no verified lookup to hands or movements, and its event durations disagree between the events table, the optical files and the cited paper, so no classifier was scored. Its optical coverage counts are published. | | [Dreem sleep cohorts · neural-network and foundation models](https://bci.report/topics/sleep-staging/#held) | 3 October 2026 | Open | The stored signal metadata says millivolts; the publisher’s own converter treats the same arrays as microvolts. Models that need absolute amplitude are not run on a guessed unit, so no such score exists. The published baselines do not depend on the unit. | | [Stieger longitudinal BCI · one-person pilot](https://bci.report/topics/model-adaptation/#cross-session) | 27 September 2026 | Open | The pipeline works on one person’s eleven sessions. With one person every score is that person’s, so none is published. | | OpenNeuro ds004902 · paired sleep comparison | 23 September 2026 | Open | Five source payloads are shorter than their own headers declare. No score of any kind. | | [Foundation models on the clinical set](https://bci.report/topics/clinical-groups/) | 23 September 2026 | Open | The source states no physical amplitude unit. Not run on an assumed one. | | L-FAME | 22 September 2026 | Open | Not included on 22 September (intake only). Since 23 September: described, not scored; its per-condition numbers withheld. | | [BNCI2015-001 · next-day adaptation](https://bci.report/topics/model-adaptation/#next-day) | 20 September 2026 | Open | Run and independently replayed on 22 September. The catalogue licence is CC BY-NC-ND and the description names no ethics approval, so no figure is published until that review is done. | Resolved holds (3) | Item | Opened | Status | Why | | --- | --- | --- | --- | | [eldBETA · one-person score](https://bci.report/topics/data-quality/#eldbeta) | 20 October 2026 | Resolved 6 November 2026 | The raw-recording route ran end to end on one participant’s 7 blocks, and the one-person score was held by the owner’s decision. All 100 participants were then scored with training-free CCA and filter-bank CCA, and the cohort result is published on the data-quality page. | | [YSU asynchronous SSVEP · consent statement](https://bci.report/topics/when-not-to-act/#non-control) | 27 September 2026 | Resolved 27 September 2026 | Scored and audited, then held until the data paper’s consent and ethics statement was read; released the same day. | | [Alpha Waves · primary consent source](https://bci.report/topics/fewer-electrodes/#posterior-subset) | 22 September 2026 | Resolved 22 September 2026 | Released the same day, after the primary report’s consent statement was read. | Each batch’s source terms, consent and privacy review: [Data use & privacy](https://bci.report/data-use/) --- Markdown copy of https://bci.report/releases/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Release archive Release log Newest first: what each reviewed batch added and held back, with the SHA-256 of every file it ships. The [holds register](https://bci.report/releases/#holds) and the [corrections](https://bci.report/releases/#corrections) are on the release log. 6 November 2026 `open-register-update-20261106` The open register in a new dated file of 550 rows: 488 dataset rows and 62 model rows. ASEEG joins this site’s dataset rows at step 4, with its result on the clinical-groups page, and eldBETA’s row is published, with its cohort result on the data-quality page. FBCCA, TS+LR, MDM, xDAWN+LDA and XDAWNCov+TS+LR are at step 5, in the core matrix; EEGSimpleConv, LaBraM at Scale and Signal-JEPA are at step 4, scored beside it; NSP is at step 3, with no checkpoint found in the sources checked; ten more models are at step 2, each with its terms by part. **Pages** [Open register · datasets](https://bci.report/datasets/register/) · [Open register · models](https://bci.report/methods/register/) · [Methods](https://bci.report/methods/) **Notes** Five dataset rows gain a finding: their dataset is named in the pretraining list of LaBraM at Scale’s Atom and Molecule collections. A row the previous file holds unchanged keeps its check date. **Notes** open-register-20261028.json and open-register-20261028.csv, and open-register-20261020.json and open-register-20261020.csv, stay served byte for byte; each is listed with its own release. **Review manifest SHA-256** `6b31ae4dc7dba62524e8f6f81eb5e8e6f25a89443f7eb93fc65aedaaf230625b` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [open-register-20261106.json](https://bci.report/data/open-register-20261106.json) | 1008.2 KiB | `ae63e86e258afdbc96e22f502c1c239bfe4af0ac2df8f978ac90c5e3267e8638` | | [open-register-20261106.csv](https://bci.report/data/open-register-20261106.csv) | 40.1 KiB | `90f407fb87285a4beb42cd2012042e37d29bfd34520da354c200915e6aff6875` | 6 November 2026 `core-matrix-update-20261106` Rows that enter the core matrix under its exact windows, people, folds and metrics, in a dated file of their own: FBCCA on SSVEP · 8 channels and SSVEP · 4 channels; TS+LR and MDM on Motor imagery & rest, Idle & command and Arithmetic & rest; XDAWNCov+TS+LR and xDAWN+LDA on P300 target ERP and Semantic target ERP; ShallowFBCSPNet and Deep4Net on seven protocols; CSP+LDA on four protocols; 30 cells in all. Each protocol page prints the rows with their recipe, role and backend, and their approved paired differences; the differences against the matrix’s one-seed EEGNet cells carry the label “against the matrix's one-seed EEGNet cell; training variance not included”. Cells that do not apply to a protocol are marked apart from cells not run. experiments.json and the eight results files keep their bytes. **Pages** [Core matrix (home page)](https://bci.report/) · [Protocols](https://bci.report/protocols/) **Notes** The seven core datasets keep their released records and terms; nothing was downloaded. pyRiemann 0.12, MNE, scikit-learn, braindecode, SciPy and NumPy are BSD-3-Clause or BSD-family; no weights are distributed. **Notes** Not published: per-person, per-fold, per-trial, per-window and per-seed values; sub-band values, fitted filters and weights; the idle mean test-window AUROC; run timings; counts of people meeting a criterion on the four-person idle protocol. **Review manifest SHA-256** `95d46dacd51131e437679bda84d034103bd1f5ff03630f1532bc9b490da34a00` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [core-matrix-20261106.json](https://bci.report/data/core-matrix-20261106.json) | 196.5 KiB | `37b41bbe9637dc15e8e442e099cc11481efb17c65bd13803c2fe94de939c0542` | | [core-matrix-20261106-mi-rest.csv](https://bci.report/data/core-matrix-20261106-mi-rest.csv) | 4.1 KiB | `49aea817f186819de7c2c0e245028484741157a352743eadd34dd09226d1787c` | | [core-matrix-20261106-idle.csv](https://bci.report/data/core-matrix-20261106-idle.csv) | 2.2 KiB | `26a084b1a720a848809d591a760845fcbcf06bd5e2ee92851f630710fee31ed2` | | [core-matrix-20261106-beta-8ch.csv](https://bci.report/data/core-matrix-20261106-beta-8ch.csv) | 3.9 KiB | `98914814219bed0d6fde404ed174d6078a7a857237d94c78034f1097f3d8298e` | | [core-matrix-20261106-beta-4ch.csv](https://bci.report/data/core-matrix-20261106-beta-4ch.csv) | 3.9 KiB | `2cbbd2c9fb34bc13050aac78e8083f2755200d4623acea5fdf91de01a97e1eac` | | [core-matrix-20261106-arithmetic-rest.csv](https://bci.report/data/core-matrix-20261106-arithmetic-rest.csv) | 6.0 KiB | `9a407a63979d162a6514b5b8c3702ae62bcab4d66ab76132a49d2d0806347d83` | | [core-matrix-20261106-p300-target.csv](https://bci.report/data/core-matrix-20261106-p300-target.csv) | 5.2 KiB | `7dd8dabd1eebaa12eb3f78b3f3e7e90c2cfc2811e78b5f94f914a0ff717b8eba` | | [core-matrix-20261106-semantic-target.csv](https://bci.report/data/core-matrix-20261106-semantic-target.csv) | 5.9 KiB | `33cede4715172ad91d08b09a902d6c1a94475f95de3afdae775efb0f2387182d` | | [core-matrix-20261106-sleep-scalp.csv](https://bci.report/data/core-matrix-20261106-sleep-scalp.csv) | 4.1 KiB | `6dd1cf11edcb7987df36680474415b8b86987f66d01826e1e9df30b76ffaffc6` | 6 November 2026 `logpower-ridge-update-20261106` BCI Report’s own log band-power ridge (5 bands, visible channels), beside the Spectral ridge row on the seven core protocols that have one. Five log band powers (1–4, 4–8, 8–13, 13–30 and 30–45 Hz) per channel on all recorded channels of the protocol, from one Hann periodogram of each published window after removing each channel's window mean and a per-window robust amplitude scaling, read by the core ridge head (alpha 100) on the published windows and participant-held-out folds. Each protocol page prints its balanced accuracy with its interval, and its paired differences from the Spectral ridge row and from one named frozen-encoder row. On the two BETA protocols it is a broad-band control stated before scoring. **Pages** [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) **Notes** Minus the Spectral ridge row, each on its own protocol: Motor imagery & rest +2.42 pp (−1.17 to +5.83), not separated; SSVEP · 8 channels −45.45 pp (−50.37 to −40.63), below; SSVEP · 4 channels −42.91 pp (−47.74 to −38.10), below; Arithmetic & rest +8.19 pp (+3.98 to +12.59), above; P300 target ERP −0.52 pp (−2.42 to +1.43), not separated; Semantic target ERP +5.08 pp (+3.06 to +7.06), above; Sleep staging −6.23 pp (−8.50 to −3.90), below. **Notes** The seven protocols keep their released data records (ds003810, ds006593 and the EESM19 scalp subset CC0-1.0; EEGMAT ODC-By 1.0; TMNRED CC BY 4.0, the stricter of its two stated terms; BETA CC BY 4.0). The comparison rows are cited by their published values under their published credits. The comparator is the project’s own code, run with numpy, scipy and scikit-learn (BSD-3-Clause); no third-party weights or code were run. **Notes** Not published: Per-window predictions, per-person scores, per-fold values and row identifiers. Features, fitted scalers and heads, the ridge alpha chosen in each fold, and the inner validation scores. The values of two internal checks, a train-selected alpha and a microvolt input view, and their per-protocol differences: the methods entry prints their two approved summary lines. An input view without window-mean removal, which was never scored. Bridge and validity-gate receipts beyond pass or fail. Compute, hosts, paths and run times. The research protocol, aggregate, audit, reviews and handoffs themselves, pinned by SHA-256 in the release manifest. **Review manifest SHA-256** `e022d08a3582b1ae825539d48fd64e44fc0aaa6f127c77868b97ee43dd4c03c8` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [logpower-ridge-update.json](https://bci.report/data/logpower-ridge-update.json) | 29.0 KiB | `152ded611dadf5d16732bd2a888afa407cafa7918f6ba6bda21579fe3b6b0ec5` | 6 November 2026 `signal-jepa-eegsimpleconv-update-20261106` Signal-JEPA as a frozen encoder with a ridge head, and EEGSimpleConv trained from scratch, on Motor imagery & rest (10 people) and Arithmetic & rest (36 people), beside the core matrix on those two protocol pages; each is compared with EEGNet on the same people. Signal-JEPA reads each window band-passed 0.5–40 Hz and resampled to 128 Hz, its pretraining format; EEGSimpleConv prints the mean of three seeds. **Pages** [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) **Notes** Motor imagery & rest: Signal-JEPA 67.25% (63.42% to 70.50%), −3.75 pp from the EEGNet cell (−7.58 to −0.17; against the matrix's one-seed EEGNet cell; training variance not included); EEGSimpleConv 59.33%, −10.14 pp from EEGNet’s three-seed mean, with no interval. Arithmetic & rest: Signal-JEPA 51.11% (48.84% to 53.47%, its interval reaching the 50% chance level), −16.53 pp from the EEGNet cell (−21.53 to −11.25; against the matrix's one-seed EEGNet cell; training variance not included); EEGSimpleConv 61.90%, −5.62 pp from EEGNet’s three-seed mean, with no interval. **Notes** ds003810 (CC0-1.0) and EEGMAT (Open Data Commons Attribution License 1.0) keep their released records; aggregate results only. Signal-JEPA’s weights are MIT as its model card at revision 51232ee0 states; its code is BSD-3-Clause (braindecode 1.5.1); the weights are not redistributed. EEGSimpleConv’s code is BSD-3-Clause (braindecode 1.5.1) and the separate author repository’s licence is unclear; it was trained from scratch, and no trained weights are shared. The model directory keeps its released notes, and a correction on record says both models were run. **Notes** Not published: Per-person, per-fold and per-window values, predictions, features, fitted heads, trained weights and participant identifiers. EEGSimpleConv’s per-seed balanced accuracies, their spread and range, and its per-seed intervals: only the mean of three seeds is published. An interval for EEGSimpleConv’s difference: the earlier study computed intervals per seed only. Measured run times, and the machines and storage paths of the runs. A partial fine-tuning arm, which was not run. The research aggregates, audits, reviews and handoffs themselves, pinned by SHA-256 in the release manifest because some carry private storage paths. **Review manifest SHA-256** `26f3fa80f776b5a15679db6ae2c079953877e457f0ad7a4c53a96a4cdd450024` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [signal-jepa-eegsimpleconv-update.json](https://bci.report/data/signal-jepa-eegsimpleconv-update.json) | 19.6 KiB | `55d25a9335d29fd1ab957cc5e3d803faf03d091ced630996758dc365af67ea00` | 6 November 2026 `foundation-checkpoints-update-20261106` Seven more encoder checkpoints in the published frozen recipe, beside the core matrix: four sizes of LaBraM at Scale’s Quark collection (B, L, H and G), the Large encoders of its Atom and Molecule collections, and REVE Small, an early checkpoint (epoch 15 of 200). On seven protocols, balanced accuracy with its participant-bootstrap interval and macro F1; on the idle protocol, trial counts with the window balanced accuracy and AUROC. The Quark ladder’s prespecified averages over six datasets, dataset-weighted: Quark-L minus Quark-B −0.93 pp (95% interval −1.75 to −0.09) and Quark-H minus Quark-L +0.03 pp (95% interval −0.90 to +0.92). **Pages** [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) **Notes** Quark-G minus Quark-H, with both encoders on the same GPU: −2.33 pp averaged over six datasets, dataset-weighted (95% interval −3.30 to −1.32); the probe's input width doubles from 800 to 1600 under one fixed penalty, so the step includes that change. **Notes** LaBraM at Scale code and weights are MIT and ungated; REVE Small is under the REVE Responsible Use License v1.0, aggregate scientific results only; no weights are distributed. **Notes** Not published: Per-person, per-fold and per-window values of any kind; features, fitted heads, idle thresholds and participant identifiers. Counts of people meeting a criterion on the idle protocol, whose cohort of four prints as a protocol fact beside the trial counts. Contrasts that were not computed: Atom-L or Molecule-L against Quark, REVE Small against REVE Base or Large, and an average of Quark-H minus Quark-B. Quark-G against Quark-B or Quark-L, any sum of ladder steps, and a GPU restatement of the Quark-L and Quark-H steps, with anything derived from them. The GPU re-run of Quark-H as cells: it is the baseline of the Quark-G step, and the Quark-H cells shown are the CPU cells. The re-extraction and attention-kernel checks as results; they are a methods note. Context values of the numeric checks: an unweighted mean over protocols, intervals from an unrelated bootstrap seed, a numeric order of the BETA participants and training-fold sizes. Weights of any kind: none were adapted, and downloaded weights are never redistributed. Timings, memory, device readings, free-space values, the compute estimate and the download time from the runs. The research records themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `b0fbab1d24e1c7effa198e921b3238932980ee4d296a9ad9f2379e1bbccdac36` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [foundation-checkpoints-update.json](https://bci.report/data/foundation-checkpoints-update.json) | 79.4 KiB | `bf32acc39f60825a271f410a91b818a75cbdbfc5bbce84fd90065b06b1f3ac04` | 6 November 2026 `random-init-controls-update-20261106` Random-initialization controls for five further checkpoints, REVE Large, ST-EEGFormer Base, ST-EEGFormer Large, SingLEM and EEGMamba, on Motor imagery & rest, Arithmetic & rest and Sleep staging: each pretrained encoder against the mean of three constructor-random initializations of the same architecture, read by the same frozen ridge head on the same people and folds as its published row, with a paired participant-bootstrap interval, under the fixed readout and, as a sensitivity table, the train-selected one. Under the fixed readout, ten of the 15 intervals lie above zero, one below it and four include zero; with no true difference, about one in twenty would exclude zero by chance. **Pages** [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) **Notes** None of the three datasets is in these four models’ published pretraining lists (checked 2026-10-04; EEGMamba’s list read from its official code). REVE Large under the REVE Responsible Use License v1.0 (reve-large @ 317531c7): aggregate scientific results only. ST-EEGFormer, SingLEM and EEGMamba under MIT. Downloaded and random weights stay local. **Notes** Not published: Per-seed values, seed ranges and seed standard deviations: only the mean of the three random initializations is published. Per-person, per-fold and per-window values; features, predictions, fitted heads and participant identifiers. Counts of people helped or harmed: none were computed. Random or adapted weights of any encoder: random states stay local, and downloaded weights are never redistributed. The alpha the train-selected readout chose in each fold. Timings, memory and device readings from the run. Bridge and gate details beyond the methods notes. The research records themselves, pinned by SHA-256 in the release manifest because several carry private storage paths. **Review manifest SHA-256** `7a08ecf864f7926d9d27acc4c3fcf3229bfc3de29eff4b36aa77225d03e94ff4` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [random-init-controls-update.json](https://bci.report/data/random-init-controls-update.json) | 41.9 KiB | `3a1eb6fa139daea44070bcdb1246164bf4ea987ad0f23e0ae53bcf81a1db75fc` | 6 November 2026 `nsp-components-update-20261106` A panel of its own on the pilot-studies page, after the counted pilots: a local test of the components of Neural State Prediction (arXiv:2609.31167v1) in three blocks. A model of 1,378,105 parameters trained on EEGMAT alone in four cells, with the target residualised or not and the context geometry topological or random; the same model never trained, at three fitting-label budgets; and the published eeg-fm-masking MAE and JEPA encoders beside a log band-power ridge (5 bands, visible channels), with channels zeroed on EEGMAT and ds003810 and the head kept or refitted. **Pages** [Pilot studies](https://bci.report/topics/pilot-studies/) **Notes** With the residualised target on, topology minus random context is +0.34 pp (descriptive interval −0.17 to +0.86), short of the 2-point margin fixed in advance. With topology context, the probe of the target’s channel falls from 77.12% without the residualised target to 8.44% with it (chance 5.26%). With all labels, each trained cell is above the same model never trained, by +1.82 pp to +2.50 pp on the mean. **Notes** EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. ds003810 keeps its released record: CC0-1.0, aggregate results only. eeg-fm-masking’s weights are CC BY 4.0 (cite arXiv:2609.33487) and its code is MIT. The NSP paper, arXiv:2609.31167v1, is CC BY-NC-ND 4.0 and is cited only; nothing from it is reproduced. The local model follows the paper’s MIT-licensed reference code. No weights are distributed, and no endorsement by the authors of either paper is implied. **Notes** Not run: The paper’s own encoder: no author checkpoint was found in the sources checked (the paper, its reference code, and model and code searches). The discriminating controls the design names: shuffled coordinates and distance-matched random context; normalisation-only, channel-only and time-only ResID; batch-partner and global-statistics controls; spectrum- and power-matched mismatch; regulariser controls; a subject-identity probe; longer or larger training and other cohorts. **Notes** The independent numeric audit replayed 60 task heads, 300 diagnostic heads, 25,920 predictions and 30,000 training steps; it did not recompute encoder outputs or optimizer updates. The first audit of this block stopped as incomplete because the auditor’s own reference rounded a training-only statistic in single precision; it is kept. A corrected auditor, checked independently, then passed all 60 jobs in double precision with the same tolerances; the producer and the aggregate are unchanged. **Notes** Not published: Per-person, per-fold, per-seed and per-window values of any kind: fold and seed-by-fold results, each job’s diagnostics and their spread, participant identifiers, the selected label rows of the fitting-label subsets, features, fitted heads, predictions, checkpoints and the sealed files. Only equal-person means, paired differences with their descriptive intervals, probe and cosine-drop means over jobs, and counts are published. Any figure or text of the NSP paper, which is cited only. Compute, hosts, storage, paths, run times and memory. Checkpoints and weights: none is distributed; each SHA-256 identifies the file used and grants nothing. **Review manifest SHA-256** `738420654f2715c0f53fe629af6fb7ccb6fde0fa6188c6340ee8d92cbbdb010d` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [nsp-components-update.json](https://bci.report/data/nsp-components-update.json) | 50.6 KiB | `8cacdeda2642721661443f246dd4cabdcdb3083b9591340e65340332d1c295da` | 6 November 2026 `electrode-shift-update-20261106` A section of the fewer-electrodes page with its own question: What if the cap sits one electrode off? A simulation on BETA (70 people, 40 targets, two-second windows): six decoders fitted at the original eight posterior sites and tested with every electrode read one position over, forward or sideways, four of them also recalibrated on the shifted cap; 12 prespecified contrasts with Bonferroni intervals, everything else labelled secondary. **Pages** [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) **Notes** Shifted minus original accuracy, decoders fitted at the original positions: one electrode forward, CCA −17.6 pp, filter-bank CCA −14.3 pp, eTRCA −28.2 pp, EEGNet −15.0 pp; one to the left, CCA −5.9 pp, filter-bank CCA −4.3 pp, eTRCA −36.9 pp, EEGNet −25.6 pp; one to the right, CCA −1.0 pp, filter-bank CCA −1.2 pp, eTRCA −42.2 pp, EEGNet −27.4 pp. Ten of the 12 read lower and two unresolved; for each, the page counts the people, of 70, losing 10 points or more. **Notes** BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. CBraMod: code MIT, weights Apache-2.0 as its official model card states; no weights are distributed. Ensemble TRCA is credited to Nakanishi et al. 2018 and filter-bank CCA to Chen et al. 2015. **Notes** Not published: Per-person, per-fold, per-window and per-seed values, including EEGNet’s per-seed means and any spread across seeds. Participant identifiers, and which channel was flagged for whom. Predictions, scores, fitted models and the sealed files. Counts of people for a 5-point change, and medians. Intervals for the analysis without the three flagged people. The 1-s spline results. Accuracies of the interpolated (spline) arms and of the shifted 1-s arms as levels; they enter only through contrasts: minus the original cap, minus the neighbour map at the same angle (the spline steps), or deployed minus recalibrated (1-s eTRCA). Storage locations, hosts, compute times and the research records themselves (pinned by SHA-256 in the manifest). **Review manifest SHA-256** `6e7e1348fccbd75940d246464ea85b570f2dd9a6a6e294e01865a62c1b1ed312` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [electrode-shift-update.json](https://bci.report/data/electrode-shift-update.json) | 92.0 KiB | `17867e64075e305d8e109dbc313e927eb752185965057633d20648b290c68007` | 6 November 2026 `eldbeta-cohort-update-20261106` eldBETA checked end to end on all 100 published participants, on the data-quality page: training-free CCA 74.38% (70.10% to 78.63%) and filter-bank CCA 87.10% (83.81% to 90.14%) mean participant accuracy, against chance of 11.11%; paired by participant, filter-bank CCA minus CCA +12.71 pp (+10.41 to +15.05). Every participant’s files pass the checks the page had reported for one participant, and the hold on eldBETA’s one-person score closes. **Pages** [Checked end to end](https://bci.report/topics/data-quality/) **Notes** Reaching the one-person binomial threshold (12 of 63 correct, one-sided 5%): 98 of 100 people under CCA and 100 under filter-bank CCA. Filter-bank CCA was higher than CCA for 84 people, equal for 11 and lower for 5. **Notes** eldBETA keeps its record’s terms, CC BY 4.0; the page prints its attribution with the modifications of this analysis. **Notes** Not published: Per-participant accuracies and correct counts, and any ordering of participants. Participant identifiers of scored or excluded people. The earlier one-participant score, which stays unrestated. Per-trial predictions, correlations, sub-band values, per-block or per-trial file-check values, windows and signals. Age, gender, BCIQ and SNR, which were never read. Per-target accuracies, confusions and ceiling counts beyond the equal group. Compute times, storage sizes, hosts and locations. The research records, pinned by SHA-256. No eldBETA data are redistributed. **Review manifest SHA-256** `61283e87459dccfb72b5d2118e693044664e3096786cccd4a50c4ab1fb15f5bc` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [eldbeta-cohort-update.json](https://bci.report/data/eldbeta-cohort-update.json) | 15.9 KiB | `2d5669fe56e8005b0d7a200cfdda8a7acfd4982db33d44ff712ee146ef41ae45` | 6 November 2026 `clinical-aseeg-update-20261106` A section of the clinical-groups page with its own question: Does resting EEG separate schizophrenia from controls better than age and sex do? ASEEG: 101 people with eyes-closed recordings, 51 with schizophrenia and 50 controls, and eyes-open recordings for 97 of the same people; stratified participant folds averaged over ten split seeds. Arms: only age and sex, EEG spectra, EEG spectra + age and sex, EEG spectra after age-and-sex residualisation, frozen CBraMod. EEG spectra minus age and sex alone: −0.05 pp (−11.51 to +11.86 pp); the interval includes zero; the difference is not an equivalence test. **Pages** [Clinical groups](https://bci.report/topics/clinical-groups/) **Notes** ASEEG keeps its record’s terms, CC BY 4.0 (Zenodo 18029536); the data descriptor is quoted only briefly, and no recording or weight is redistributed. **Notes** Not published: Participant rows of any kind: file codes, ages, sexes, diagnoses, fold assignments, features, out-of-fold probabilities and fitted heads. Per-fold and per-split-seed values, the split seeds themselves, and seed ranges. Per-group demographics: age means, medians and ranges, sex counts, and any figure derived from them, the data descriptor’s own included. Confusion counts: seed-averaged values are not whole people. Recording lengths of individual files. Bridge numbers beyond pass or fail and their largest differences, and smoke-test numbers. EEG signals, the descriptor’s figures or text beyond short quotes, and model weights. Runtimes, hosts and storage locations; the research records themselves are pinned by SHA-256 in the release manifest. **Review manifest SHA-256** `0e244c2c30a799dcfdf9a4429e0f6311a2a1a6f52fe663f6f25da439740f809e` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [clinical-aseeg-update.json](https://bci.report/data/clinical-aseeg-update.json) | 23.4 KiB | `f0dacf3df3ac2391e043bb9d2a17907a92f93b73b208d36272257d941362c5ea` | 28 October 2026 `open-register-update-20261028` The open register in a new dated file of 535 rows: 487 dataset rows and 48 model rows. The four datasets checked end to end on the data-quality page become this site’s rows at step 4 of the dataset ladder, each with its MOABB record folded in: Kim beta-range SSVEP and MIND with their results published, and Damm finger motor imagery and NeBULA with their scores held. Three models scored on the pilot-studies page enter the model ladder at step 4: NeuroRVQ, EEG-Captioner and AXON. Five more models are identified from their papers at step 1, each with its hold: BandVQ, EEG-PRIME, FAME, STST-JEPA and USE-FM. open-register-20261020.json stays served byte for byte. **Pages** [Open register · datasets](https://bci.report/datasets/register/) · [Open register · models](https://bci.report/methods/register/) · [Methods](https://bci.report/methods/) **Notes** Kim beta-range SSVEP: its row links its dataset page and its case on the data-quality page, and holds models whose output depends on amplitude, in place of the onset-timing hold its MOABB record carried. MIND: both licences side by side (CC BY 4.0, CC BY-NC 4.0), with original terms and pinned version as gaps at step 3. Damm finger motor imagery: rights review and consent or ethics statement as gaps at step 3. NeBULA: its Figshare record at its only version, with no gap at step 3. **Notes** NeuroRVQ, EEG-Captioner and AXON: each links its panel on the pilot-studies page and has no method page; pretraining exposure is unknown for all three. Terms by part: NeuroRVQ: code CC BY-NC 4.0, weights CC BY-NC 4.0; EEG-Captioner: code MIT, weights Apache-2.0; AXON: module code Apache-2.0, project code BSD-3-Clause, weights Apache-2.0. BandVQ, EEG-PRIME, FAME, STST-JEPA and USE-FM: each links its paper; EEG-PRIME is held because its code repository could not be reached, and the other four because no official release of code or weights was found. Not published: the register use statements drafted for some rows, which stay internal, and third-party conversions of NeuroRVQ’s checkpoint. **Review manifest SHA-256** `c187617c0529598213cadca9d24d498dc786216599a69a118ff6ba944af2d282` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [open-register-20261028.json](https://bci.report/data/open-register-20261028.json) | 969.5 KiB | `0df2dc782a1bcbb6844cba8f66185f2cbd3d5e98fc9d8ca580672585d0d78bff` | | [open-register-20261028.csv](https://bci.report/data/open-register-20261028.csv) | 40.0 KiB | `c499d614dd3999c7679cc2f7e5701e0e93f20fdefebc7c23707b49ec632432c5` | 28 October 2026 `ssvep-followups-update-20261028` Seven more follow-ups on SSVEP frequencies a decoder never trained on, as cards 8 to 14 of the same question page, each read from its own audited aggregate: the cross-spectral round’s results, which take the place of card 8’s status entry; eigenvalue pooling; harmonic pooling, on BETA and on Wearable-102; filter-bank CCA with shorter windows and fewer electrodes; the gap between harmonics; a fixed decay over harmonics; and rising harmonic weights. All seven use route 3’s BETA cohort (70 people), and four of them also run on Wearable-102’s dry and wet headbands (102 people). Seen and held-out frequencies are separate tasks, reported side by side, and each set-up sits beside its untrained control. Each card carries its own interval method: card 12’s primary family is Bonferroni-adjusted over 9 comparisons, card 13’s margin test uses a 90% interval, and no other interval is adjusted. Card 11 adds CPU timings: warm, single-thread measurements of each component, never summed, and no sensor-to-decision latency. **Pages** [New SSVEP frequencies](https://bci.report/topics/new-ssvep-frequencies/) **Notes** On the held-out 8-way task with 2-second windows and 8 electrodes, filter-bank CCA with no training reached 90.21% against 80.90% for CCA on the same windows (+9.31 points, +6.84 to +11.71; 70 people). The learned eigenvalue head differs from the harmonic sum, which needs no training, by +0.07 points (−0.45 to +0.62). Route 3’s questions-in-language page runs another filter-bank recipe on the same task, with 5 Butterworth sub-bands. **Notes** BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. Wearable-102 keeps its released record: CC BY 4.0, cohort aggregates only. Filter-bank CCA is credited to Chen et al. 2015; the code that ran is the project’s own SciPy realisation, and no code or weights are distributed. This file’s card 8 replaces the status entry that new-ssvep-frequencies-update.json carries for it; that file keeps its bytes. **Notes** Not published: per-person, per-fold, per-rotation, per-seed, per-window and per-frequency values of any kind, predictions, scores, features, fitted heads and the sealed run files; the cross-spectral round’s per-seed, canary and random-pairing accuracies; the fixed-decay retrospective ratio table; all-trial decision counts beyond the harmonic-gap shares, and that card’s Delta means and distractor margins; run times of whole evaluations and audits, memory, CPU-time readings and every host, processor, system and runtime label; the comment thread’s predictions on these cards; the commenter’s name; and the handoffs, protocols, freezes, receipts, reviews and approvals themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `481fefa22dac4febdce98935fe31c432ff8d438d3217dcc46fbd765275f005a3` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [ssvep-followups-update.json](https://bci.report/data/ssvep-followups-update.json) | 317.0 KiB | `e147f2f0bb0babf103a138aeab9cf9a59c0bd153bad477cc0a193907253ddbb5` | 28 October 2026 `neurorvq-update-20261028` A separate panel on the pilot-studies page, outside the pilot count and every leaderboard: NeuroRVQ’s released backbone, kept frozen and read by one fixed ridge head, on three existing tasks: EEGMAT’s arithmetic and rest (36 people, 2,160 two-second windows), motor imagery against rest on ds003810 (10 people, 1,200 windows) and P300 target against non-target on ds006593 (21 people, 2,520 one-second epochs), 5 participant-held-out folds each, nothing tuned. Each task sets the backbone’s 800 pooled values beside simple features of the same windows, with participant-mean balanced accuracy and macro-F1 for each, and one paired difference with its participant-bootstrap interval. Pointers on the three protocol pages lead to it and carry no figure; the core matrix keeps its rows. **Pages** [Pilot studies](https://bci.report/topics/pilot-studies/) **Notes** NeuroRVQ minus its comparator, in balanced-accuracy points: on EEGMAT +1.53 (−3.52 to +6.67) against band power: the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent; on motor imagery −15.25 (−20.17 to −10.25) and on P300 −3.57 (−5.56 to −1.51) against temporal means, both below zero. Pretraining exposure is unknown: none of the three datasets is named in the paper’s list of 13 pretraining corpora, 1 of them proprietary. There is no random-initialisation arm, so the panel says nothing about pretraining. The authors’ inference wrapper does not restore strictly (8 normalisation tensors missing), so what ran is the released backbone with parameter-free branch normalisation and mean-token pooling. **Notes** NeuroRVQ’s code and weights are CC BY-NC 4.0, the licence file of both repositories at the pinned revisions; only aggregate figures are published, no weights are distributed, and no endorsement by NeuroRVQ’s authors is implied. EEGMAT, ds003810 and ds006593 keep their released records. Not run: semantic-target (the original calibration evidence is still missing); beta-4ch (physical units and reference are unresolved); beta-8ch (physical units are unresolved, and PO5 and PO6 are missing from the model’s channel list); idle (its inputs are missing, and it needs its own calibration, threshold and split protocol); sleep-scalp (deferred to a separate study). Not published: per-person, per-fold and per-window values of any kind, fold assignments, identifiers, labels, extracted values, fitted heads, predictions and the sealed files; how many people each arm scored above 50%; per-arm intervals and macro-F1 differences, none of which was computed; measured compute; the checkpoint and the authors’ source; third-party conversions of the checkpoint and the licence tags they carry; and the handoffs, protocols, configurations, receipts and reviews themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `876311164c20374afa7c085710df03877a31d778e3e88e33028888127d898391` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [neurorvq-update.json](https://bci.report/data/neurorvq-update.json) | 31.4 KiB | `6ae57314d029016cc5934d7c24d8330f33b9afeaf82ff4380f9c4fae38e5f24d` | 28 October 2026 `eegmat-ten-second-update-20261028` Another separate panel on the pilot-studies page, outside the pilot count and every leaderboard: two frozen encoders on ten-second EEGMAT windows of arithmetic and rest (36 people, 72 recordings, 432 windows, 5 subject-disjoint folds), each read by one fixed ridge head and set against Welch band power on the same windows, in a subsection of its own: EEG-Captioner’s encoder (256 dimensions; only the encoder ran, with no caption) and AXON (512 dimensions). Participant-mean balanced accuracy and a paired difference with its participant-bootstrap interval for each; beside every score, 0 of 432 windows meet EEG-Captioner’s source annotation-QC rules. No difference between the two encoders is computed, and the panel is kept apart from the two-second results on this site. **Pages** [Pilot studies](https://bci.report/topics/pilot-studies/) **Notes** EEG-Captioner’s encoder reached 65.74% against 67.59% for Welch band power fitted in the same run (−1.85 points, −9.03 to +4.86). AXON reached 70.14% against the same band-power predictions, reused without refitting and matched for neither normalisation nor parameter count (+2.55 points, −5.79 to +10.42). Both intervals span zero. AXON’s own numerical input check passed 432 of 432 windows, a different check from the source QC, which flags whole recordings (72 of 72) and does not say that each window has an artifact. AXON ran after the EEG-Captioner and Welch outcomes were known; rest is recorded before arithmetic; pretraining exposure is unknown for both encoders. **Notes** EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. EEG-Captioner’s source code is MIT. EEG-Captioner’s encoder weights are Apache-2.0 as its Hugging Face weight metadata states. AXON’s checkpoint is Apache-2.0 as its Hugging Face licence tag states; its Braindecode module carries per-file Apache-2.0 terms within a BSD-3-Clause project. No weights are distributed, and no endorsement by either model’s authors is implied. Not published: per-person, per-fold and per-window values of any kind, fold assignments, identifiers, features, fitted heads, predictions and the sealed files; confusion matrices and pooled macro-F1; the source-QC breakdown beyond its totals; forward counts, head-fit counts, CPU and wall time and memory; a difference between the two encoders; AXON’s paper reference, until it is checked; the checkpoints and weights, whose SHA-256 identifies each file and grants nothing; and the handoffs, protocols, approvals, reviews, audits and receipts themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `3f22f09009af5b2dd989504f21c6776b1c2b38b9a559ed4a6d9257219918cd0f` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [eegmat-ten-second-update.json](https://bci.report/data/eegmat-ten-second-update.json) | 22.9 KiB | `b26e1934e8569ac8de8f51f815dad1dcc74a0734f9090b5e2a70572ea1cc5fe0` | 28 October 2026 `dataset-checks-update-20261028` Four more public EEG datasets checked end to end on the data-quality page, each read from its own files. Kim beta-range SSVEP is a worked case: training-free CCA on all 40 participants (9,600 trials, 40 targets from 14.0 to 21.8 Hz), with a dataset page of its own. MIND is a conditional case on record IDs: CSP + LDA on two stored-code pairs beside an ordinal lookup and a majority control, with record-ID resampling ranges and the facts of its label schedules. Damm finger motor imagery and NeBULA are one-person status cases: design, counts and file facts, with every score held. The one-person rule is stated once on the page, before eldBETA, Damm and NeBULA. **Pages** [Checked end to end](https://bci.report/topics/data-quality/) **Notes** Kim: mean participant accuracy 53.40% (45.89% to 61.01%), against chance of one in 40; the amplitudes are as stored, and models whose output depends on amplitude are held. MIND, primary panel of 28 record IDs and 3,360 trials: CSP + LDA 56.96% on stored codes 4/5 and 51.85% on 6/7, against 50.00% for the majority control; the ordinal lookup sits below the majority control because each pair follows one of only two complementary 20-trial label schedules. Record IDs are not verified people. Damm: 150 delivered cues over five fingers and 62 scalp channels. NeBULA: 89 delivered trials, 127 EEG channels and 11 EMG rows where the metadata states 16. **Notes** Licences, from each record: Kim CC BY 4.0 (Figshare 28806815, version 3, which the research used; the paper cites version 2, also CC BY 4.0); MIND CC BY 4.0 and CC BY-NC 4.0 side by side, as the paper and the archive’s dataset description state the first and the MOABB adapter and the ScienceDB record’s metadata the second, with ScienceDB’s terms-of-use page not read; Damm CC0 as its dataset description states, while its README cites a LICENSE file the tree does not hold; NeBULA CC BY 4.0 (Figshare 27301629), the paper’s own licence covering the article only. Kim’s, MIND’s and NeBULA’s papers state ethics approval and consent, MIND’s without naming its committee; Damm’s dataset description names its ethics approval, and no consent statement was found. Not published: values per person, per record ID, per fold, per run or per trial, identifiers, features, fitted models, predictions and raw signals; every Damm and NeBULA score and any comparison of NeBULA’s order diagnostic with its EEG or EMG arms; MIND’s pooled metrics, confusion matrices, secondary schedule census and model-fit counts; Kim’s per-participant accuracies; compute, hosts and storage locations; and the research aggregates, audits, reviews, protocols and handoffs themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `fce3cacdbf811855fd224a4704c392becf8829673b6967826a558ece8cbe6202` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [dataset-checks-update.json](https://bci.report/data/dataset-checks-update.json) | 52.2 KiB | `80c504fed02f4aadad33ee5c3151a54a16ac3f4b09ddae3f1805c9073184104d` | 20 October 2026 `open-register-update-20261020` The open register: every EEG dataset we know of and every EEG model and paper we track, each with the furthest step we have checked and what that step showed, in 527 rows (487 dataset rows and 40 model rows). Two ladders with their own step names: a dataset is listed, read, traced, run here and in the core matrix; a model is identified, has its terms recorded and its files checked, is scored here and in the core matrix. Each row keeps the evidence for every step as pass, gap or not applicable, with its source, so its step is the furthest one whose evidence is on file and its gaps stay visible. The dataset rows come in five groups: uploads on Hugging Face that hold brain-signal data, from a frozen snapshot of 28 search terms and 20 tag filters read on 19 October 2026 at 14:23 UTC (1,922 listings, 737 of them EEGDash mirrors, and 368 other uploads in scope); OpenNeuro’s mirrors on EEGDash, as one family row; this site’s own datasets; the catalog of MOABB release 1.7.2, one row per source record (206 loaders over 158 records, 18 of them merged into this site’s rows); and corpora that require an agreement or request, with their access terms as captured from the source. Anything can be listed here; the score tables keep their own rules. **Pages** [Open register · datasets](https://bci.report/datasets/register/) · [Open register · models](https://bci.report/methods/register/) · [Datasets](https://bci.report/datasets/) · [Methods](https://bci.report/methods/) **Notes** The rules, stated once on the register: gated data, or data that needs a signed or requested agreement, stays at step 1 unless we accept its terms; suspected re-uploads stay at step 1 whatever their gating, compared by metadata only, with none of their files downloaded; a finding about a re-upload is stated only when an independent check confirms it and a captured page of the original terms backs it, and then in factual words: that an upload matches another source’s metadata, or that its declared terms differ from the original terms; rows state facts; and every row has a correction route, a mail that names the row’s id. The CSV carries no findings; they are in the JSON. Repository ids and usernames appear in a row’s text, never in a link or an anchor; an uploader’s series of repositories is one family row; card text, download counts and modification dates are not shown. **Notes** The register now lists sources that give a reader nothing to run, such as L-FAME (CC BY-NC 4.0), because it records what was checked. Each later register version is a new dated file and release; this file stays served and listed. **Review manifest SHA-256** `e72bbccb44930f7d1072d4d9fcc39c773f558e0b37d778a0092faa3019bfc622` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [open-register-20261020.json](https://bci.report/data/open-register-20261020.json) | 939.1 KiB | `c131c470df881e51389aca58d17ec3c0ac37eb4244c082ff7ffe20582b77d558` | | [open-register-20261020.csv](https://bci.report/data/open-register-20261020.csv) | 39.4 KiB | `aef5cffcac9fa23b14a0276ee0a8d9cb79aecf1d0f81eb95613f7d7874b7a9fe` | 20 October 2026 `new-ssvep-frequencies-update-20261020` Route 3’s follow-ups on SSVEP frequencies a decoder never trained on, on a new question page: seven audited studies that answer questions from a reader’s comments on Hugging Face. Six reuse route 3’s BETA cohort and splits (70 people, 7 participant folds by 5 frequency rotations, 32 seen and 8 held-out frequencies; trained heads over 3 seeds), each testing one idea: asking with the reference spectrum; the head, the spectrum or the question; a random delay on a phase-carrying query; a random phase ramp with magnitude queries; where CCA’s lead comes from; and a bounded learned correction on top of CCA. The seventh runs the correction’s models unchanged on Wearable-102’s dry and wet headbands (102 people). Each card gives its question, what changed, the matched control, the result and its limit; every interval is its own study’s, and no multiplicity correction was applied. A status card for the cross-spectral round carries no figures; that round ships as its own release. **Pages** [New SSVEP frequencies](https://bci.report/topics/new-ssvep-frequencies/) **Notes** On frequencies left out of training, temporal CCA with a small learned correction reached 86.15% against 80.90% for CCA alone, choosing among 8 frequencies (+5.24 points, 3.91 to 6.62; 70 people). Run unchanged on Wearable-102, the same models changed CCA’s score by +0.07 points on dry electrodes (−0.71 to 0.91) and −0.61 on wet (−1.48 to 0.23). Learned heads that read reference spectra or stimulus phase stayed far below CCA. **Notes** BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. Wearable-102 keeps its released record: CC BY 4.0, cohort aggregates only. The frozen CBraMod features are the ones route 3 used; no weights are distributed. The stimulus phases of one query come from the BETA authors’ corrigendum, read at a pinned revision of the Hugging Face mirror Bingchuan/BETA. The Wearable-102 documentation cannot rule out that some of its participants also appear in BETA. **Notes** Not published: per-person, per-fold, per-rotation, per-seed, per-window and per-frequency values, predictions, scores, features, fitted heads and the sealed run files; the adapter’s per-seed accuracies and the transitions of the cropped-match correction; the Wearable-102 contrast against the cropped-spectrum check; the commenter’s name; measured compute; and the handoffs, protocols, freezes, receipts and reviews themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `54cc10869787ca8af1b9bff5e7306eadc5bb1e5c277f27ae95c8f1127702981a` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [new-ssvep-frequencies-update.json](https://bci.report/data/new-ssvep-frequencies-update.json) | 212.3 KiB | `197028bdc4dadd9713f6d408bb64204147fdfe39a686e7e119e4988aaf863ee0` | 20 October 2026 `data-quality-update-20261020` A new question page: what do the files show once a public EEG dataset is checked end to end? Three public datasets, each read from its own files. Shin2022 is the worked case: a participant-held-out diagnostic over 10 people with three arms (a participant-weighted training prior, a normalised spectral ridge on the first second of cursor feedback, and a retrospective control that reads only the feedback duration), scored on the signal-valid windows that its coverage chain counts from the runs in the release. eldBETA and ZuCo 2.0 are status cases with no score: one participant’s eldBETA recordings, as counts, and a timing hold on one converted ZuCo 2.0 run. **Pages** [Checked end to end](https://bci.report/topics/data-quality/) **Notes** Shin2022: in the release, 9 pairs of runs are byte-identical, so the files hold 101 distinct runs, and one incomplete run is held whole. On the signal-valid windows the normalised spectral ridge reached 52.59% (51.51–53.66) participant-mean balanced accuracy and the duration control 52.33% (50.87–54.02), against 50.00% for the prior. The scores cover the signal-valid windows only: feedback lasting at least one second, and a first second that is finite with positive 1–40 Hz power on all 16 channels. The label is the instructed target in closed-loop cursor recordings, which carry visual, controller and behaviour confounds. The duration control reads the whole feedback duration, known only after the trial ends, so it checks a confound; the spectral arm sits close to it, and no test between arms was run. Ten participants, one recording day each; the intervals are descriptive. The catalog’s six-second trial length is the feedback cap (MaxFeedbackDuration) that every run’s header sets. **Notes** eldBETA: the raw-recording route runs end to end on one participant’s 7 blocks: all 63 trials pass the reader and the quality checks, and none is dropped. The continuous recordings hold 248,320 non-finite values outside the analysed windows, left as recorded. With one person every score is that person’s, so the accuracy is held, by the owner’s decision. The header calibration was applied, and the derived MATLAB epochs were not used. ZuCo 2.0: in one converted run on NEMAR, some fixation times fall outside the recording, so no EEG–text score is computed until the converter’s timing rules and a source-checksummed fixture are available. **Notes** Licences, from each record: Shin2022 CC BY 4.0 (Figshare 20383716), eldBETA CC BY 4.0 (Figshare 18032669), ZuCo 2.0 CC-BY-4.0 as its NEMAR dataset description states, with the original terms on OSF and the terms of the quoted text excerpts unverified. Shin2022’s paper names its approving board and states written informed consent; eldBETA’s data descriptor gives its board’s approval number and states written consent; for ZuCo 2.0 no consent or ethics statement was read in this release. Not published: participant rows of any kind and raw signals; Shin2022’s per-participant figures, feedback-duration distribution and per-channel census counts; eldBETA’s one-participant accuracy and its comparator; ZuCo 2.0’s converted-run code, timing values and marker counts; compute; and the research exports, audits, reviews, protocols and handoffs themselves, pinned by hash in the release manifest because several carry private storage paths. MOABB’s catalog counts for Shin2022 are findings on its row in the open register. **Review manifest SHA-256** `57d38a4f89486fec7edce50825d7ac977af25cfae035bb9bda9a6393aca4e005` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [data-quality-update.json](https://bci.report/data/data-quality-update.json) | 27.9 KiB | `fbfb8fb59dc80397d0c4331a3067cf640b27522af6687dd5b3132c8f4e928bda` | 20 October 2026 `clinical-iesseeg-update-20261020` A new section on the clinical-groups page, with its own question: do EEG case-control results survive age and recording-setting controls? IESSEEG, a public infant EEG dataset (100 people: 50 cases of infantile epileptic spasms syndrome and 50 controls, one awake minute each), on the publisher’s 5 participant-disjoint folds: six arms with and without EEG, including an age-only control, a recording-settings-only control and EEG after age is regressed out inside each training fold, and four arms that read frozen NeuroLM features with the same probes. Out-of-fold point estimates of balanced accuracy and AUROC, both shown for every arm, for everyone and for the 68 people aged 2 or under. The page’s Parkinson’s question, short answer and card keep their wording, and IESSEEG adds no dataset page or count. **Pages** [Clinical groups](https://bci.report/topics/clinical-groups/) **Notes** Point estimates out of fold; no interval or test was specified. One awake recording per person, first 60 s. Age and recording settings differ between groups; regressing out age is a sensitivity analysis. The probes and every score were reproduced independently; the NeuroLM features themselves were not re-extracted by a second implementation. NeuroLM’s preprocessing differs from the spectral pipeline: the two 2,000 Hz recordings are resampled to 200 Hz, then a 0.1–75 Hz band-pass, 50 and 60 Hz notches and division by 100. A difference between the two pipelines therefore mixes preprocessing with architecture. The checkpoint’s training stage and its prior exposure to these participants are unknown. **Notes** IESSEEG: CC BY 4.0 for the dataset and annotations, from the release card and LICENSE at revision 1c32ee3a of Capur/IESSEEG, cited by the title the card requests: “IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome”. The release card states institutional review board approval of the retrospective study, with waivers of informed consent and HIPAA research authorization, and asks that no one attempt to re-identify participants. The publication boundary is the one the Parkinson’s comparison uses: aggregate scores only, with no participant rows, predictions, ages or per-group age and sex composition; the age ≤2 column gives the band’s size and its case and control counts. The card declares that the 400 case segments were previously analysed by Rajaraman et al. (2024); the participant-level crosswalk to that analysis is unresolved. NeuroLM-B is used frozen and identified by its checkpoint’s SHA-256; no weights are distributed. Not published: participant rows of any kind, per-fold metrics, per-group age and sex composition, intervals and tests, signals, code, the checkpoint and its features, and runtimes and storage locations. **Review manifest SHA-256** `67418da30bd51f1c89985c845f88e6dcc2d8b875ab6da04b7dde4d9b179ce7ac` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [clinical-iesseeg-update.json](https://bci.report/data/clinical-iesseeg-update.json) | 27.2 KiB | `c2302fe6945276de8f9183340c165245bc4c5097c9ebe60fa7ac389725b2427d` | 20 October 2026 `neurolm-eegmat-update-20261020` A separate panel on the pilot-studies page, outside the pilot count and every leaderboard: NeuroLM-B kept frozen on EEGMAT’s arithmetic and rest (36 people, 2,160 two-second windows, whole participants held out in 5 folds), with one fixed ridge head fitted on each of three representations of the same windows: NeuroLM’s tokenizer features, its GPT hidden states and band power. Equal-person balanced accuracy and macro-F1 for each, and three paired differences with paired participant-bootstrap intervals. A pointer on the Arithmetic & rest protocol page leads to it; the core matrix keeps its rows. **Pages** [Pilot studies](https://bci.report/topics/pilot-studies/) **Notes** Tokenizer features minus band power is +0.60 points (paired 95% participant-bootstrap interval −5.32 to +6.44): the interval spans zero, so neither is shown to be better, and the two are not shown to be equivalent. Pretraining exposure is unknown: EEGMAT is NeuroLM’s Workload task, one of the 6 instruction tasks in its pinned code, and this checkpoint’s training stage is unresolved. The pilot has its own adapter and band-power features, so its scores sit in their own panel; rest is recorded before the task, so recording order moves with the label. **Notes** EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. NeuroLM’s code is MIT at the pinned commit. The NeuroLM-B checkpoint is CC BY 4.0 as the licence tag of its Hugging Face repository states, and its 30-byte model card adds no terms. No weights are distributed, and no endorsement by NeuroLM’s authors is implied. Not published: per-person, per-fold and per-window values of any kind, demographics, measured compute, the checkpoint and the NeuroLM source, and the handoff, protocol, producer, receipts and reviews themselves, pinned by hash in the release manifest because several carry private storage paths. **Review manifest SHA-256** `41c6c94772c50dcafb9624ad2d5ae20274a3f5b8d4617ef14a7e38105a45861e` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [neurolm-eegmat-update.json](https://bci.report/data/neurolm-eegmat-update.json) | 19.2 KiB | `4e7001b2b69ea2578d1db5c9617dc1b99e8f3985f13b2e2fb1b11f32024d6a55` | 13 October 2026 `eegpt-core-update-20261013` EEGPT on the core matrix: its eight core-matrix cells, computed in the same runs as the published LaBraM rows and, except on idle, the CBraMod rows, with only the encoder swapped, published as computed, without a re-run, as a frozen row beside the core matrix, where the further foundation encoders sit. On seven protocols, balanced accuracy with its participant-bootstrap interval, macro F1 and the chance level; on the idle protocol, counts (idle false activations, commands detected within 3 s, people who always abstained) with the window balanced accuracy and AUROC, and no interval. Beside the published LaBraM and CBraMod rows, EEGPT’s pretraining-attribution rows on mi-rest and EEGMAT: the pretrained encoder against the mean of three random initialisations, under a fixed and a train-selected readout, with paired intervals. Grouped by family, never ranked: EEGPT is above or below a published row only where their 95% intervals do not overlap, and the gap between the intervals is given in percentage points. Each cell says whether its dataset is in EEGPT’s published pretraining list, checked on 12 October 2026 with the source (none of the seven core datasets is), and each footnote sits beside the cells it bears on. **Pages** [Core matrix (home page)](https://bci.report/) · [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) **Notes** Several gaps are narrow. On sleep staging EEGPT’s interval lies above every core-matrix row’s, the nearest by 0.10 pp; on P300 it lies above LaBraM’s by 0.28 pp, on the semantic target above the spectral ridge’s by 0.78 pp, and on BETA with four channels above LaBraM’s by 0.90 pp. On BETA, EEGPT lies below standard CCA, which needs no training, at both channel counts. The idle row comes from the same run as the published LaBraM idle row; the published CBraMod idle row comes from another run, so a comparison with it is across runs. Pretraining attribution: on both datasets and under both readouts the pretrained encoder scored above the mean of its random initialisations, with paired intervals above zero; a frozen-encoder result conditional on these splits, readouts and initialisations, not a model ranking. **Notes** Owner decisions of 12 and 13 October 2026: publish EEGPT’s eight core-matrix cells as computed, without a re-run, as a frozen row beside the core matrix, and its pretraining-attribution rows beside the published LaBraM and CBraMod rows. The rights reasoning is the pilot comparison’s of 12 October: independently computed aggregate results are a separate output from the weights; no publisher prohibition was found; the checkpoint’s licence is unresolved; no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. No author endorsement is implied. Four of the eight cells (mi-rest, EEGMAT and both BETA protocols) and EEGPT’s attribution controls have been in audit records in the public repository since 20 September 2026; the public attribution audit still labels the four EEGPT contrasts “hold-unresolved-weight-license”, as it did then, and keeps its bytes. With this release they are published. **Notes** How EEGPT differs from EEG-DINO: EEG-DINO stays catalogue-only by the owner’s decision of 4 October 2026 — its upstream weights carry no licence, and it was neither downloaded nor run. EEGPT was evaluated in the same runs as the published rows and was held only because its checkpoint’s licence could not be read, a hold the reassessment of 12 October found to be this project’s own choice. EEG-DINO is not re-reviewed by this decision. **Notes** Not published: per-trial, per-person, per-fold and per-class values, features, fitted heads, idle thresholds and calibration grids, and how many people had a negative attribution difference; the per-seed scores of the three random initialisations (their mean is published, as for LaBraM and CBraMod); BETA’s duration-group summaries; timings, devices and memory; EEGPT results outside the eight core protocols; the checkpoint and its tensor conversion, whose SHA-256 identifies them and grants nothing; and the run summaries, protocols, independent audits and verification records themselves, pinned by hash in the release manifest because they carry private storage paths. Nothing was re-run: experiments.json, deployment-topics.json and the further-encoder files keep their bytes. **Review manifest SHA-256** `21e27baaebe9af009a9555a5b1a64e367246b14df00413ab0d9159c183b60ffc` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [eegpt-core-update.json](https://bci.report/data/eegpt-core-update.json) | 74.1 KiB | `8563862b961a226636d1349ec62f5bfd4fe922bda326c27fbcf7557e2d277a26` | 12 October 2026 `pilot-studies-update-20261012` Small pilots on four sources, each from its own approved aggregate export, all kept outside every leaderboard and model ranking. The Orthosis error-detection dataset (8 people, 83 runs, 2,407 recorded device events: one cohort, counted once) in three results: a 64-channel classical EEG baseline, the same baseline on the 58 channels EEGPT supports, and frozen EEGPT with the same fixed supervised probe against those matched classical features. BINGO’s Greek and English inner speech (17 source-defined subject fields): a fixed spectral baseline across people and across languages, a negative result near its 15-class prior. A mask-aware EEG analysis of an imagined-handwriting competition dataset (5 people, four released coded labels): with channels that sit flat at a digital rail masked, relative EEG is slightly above the training-majority control and above the missingness-mask and mask-plus-time controls, and not separated from elapsed time and trial order; a data-quality and methods case study, not handwriting recognition. REFED (two nominal participant folders, 30 stored clips): a computational diagnostic in which simple source-only controls have smaller errors than EEG band power. Nothing is pooled across datasets, configurations or conditions. **Pages** [Pilot studies](https://bci.report/topics/pilot-studies/) · [Methods](https://bci.report/methods/) **Notes** EEGPT: the comparison is published from its approved export, after an independent check of that exact file. The earlier hold on EEGPT aggregate scores was this project’s own conservative choice, not a publisher’s rule; this pilot comparison is the first EEGPT result on the site’s pages. The core-matrix EEGPT scores are not on the site’s pages, but some of them (mi-rest, EEGMAT and both BETA protocols, with EEGPT’s pretraining-attribution controls on mi-rest and EEGMAT) have been in audit records in the public repository since 20 September 2026. The checkpoint’s licence remains unresolved, and no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. The paired interval spans zero, so neither arm is shown to be better, and the two are not shown to be equivalent. The probe is trained on the other seven people’s labels (supervised transfer, not zero-shot), and overlap with EEGPT’s pretraining data has not been excluded. Corrected the same day: this note first said the pilot comparison was the only EEGPT figure published. The core-matrix cells followed on 13 October 2026, in eegpt-core-update-20261013. **Notes** BINGO: the source states no licence. Consent is as the project’s public deliverable D3.1 (version 3.0) reports it: signed consent forms, stated for the project’s NATO-alphabet recordings in the same CERTH laboratory; the cross-language dataset’s own documentation has no ethics or consent statement, and no ethics-approval reference is published. The aggregates are published on that basis by the owner’s decision. The project documents artifact cleaning (ASR and FORCe) on the continuous recordings before trials were cut; that statement is printed beside the result, attributed to D3.1 (version 2.0), and was not re-verified against the file. The project’s own result on its NATO-alphabet dataset is printed as reported by the project, not run here. **Notes** Handwriting: published by the owner’s decision with its rights gap stated. No ethics-approval or informed-consent statement was found, on the dataset card or in a participating team’s challenge paper, and the organisers’ paper is listed as forthcoming; to be rechecked when it appears. The earlier analysis of the same five people is printed as context, not as a second cohort. **Notes** Not published: per-person, per-field, per-folder, per-clip and per-record values; the Orthosis response-clean subset’s scores (its counts are published); individual permutation and label-shuffle replicates (their mean and observed range are published); confusion matrices; the handwriting mask-only control’s interval, because with five people its bounds would show individual people’s scores; and the handoffs, approval records and independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. Status only, with no figure: Orthosis EMG, TinyMyo and EEG–EMG fusion; handwriting fNIRS and fusion; REFED physiological and fusion evaluation; and a meditation dataset, in which RR means heartbeat RR intervals, not respiratory rate. **Review manifest SHA-256** `89dcd9295ad0a85abea68deb180fcbed25165d4f523d5e02b402937eeb512c29` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [pilot-studies-update.json](https://bci.report/data/pilot-studies-update.json) | 94.6 KiB | `c93a344b7111b2335565b4a2819c0f4be6a93ba8d6970a8fd4fcb0cf9caf2bca` | 12 October 2026 `beyond-scalp-update-20261012` Locally executed analyses beyond scalp EEG, each its own species, signal and metric: none is compared with scalp EEG, none joins the core matrix, and they share no ranking. Human fNIRS (Bak et al., 30 people): executed right-hand, left-hand and foot tapping classified by a fixed shrinkage LDA in a person left out of training, against a training-prior comparator. One macaque’s motor-cortex spiking (MC_RTT, prepared for the Neural Latents Benchmark; one session): x/y finger velocity decoded by a fixed ridge on held-out trials of the same session, against the training mean, on five metrics, including the one on which the ridge is worse. A numerical coverage check of human fNIRS recordings (ds007420, 12 people, 60 recordings): counts only, no decoding. Two sources are prepared and not scored: FALCON H2 human intracortical handwriting and Rabut et al. human functional ultrasound. **Pages** [Beyond scalp EEG](https://bci.report/topics/beyond-scalp-eeg/) **Notes** Bak: the participant-macro balanced accuracy is well above the training prior, with a descriptive interval that resamples fixed participant scores (not a refit bootstrap). Executed movement, not motor imagery; offline over a 15-second window, and systemic or motion effects are not excluded. MC_RTT: not an official Neural Latents Benchmark score; with one animal every figure is that animal’s, and no interval is reported. ds007420: coverage counts say nothing about signal quality, and no classifier score exists, because its task codes have no verified lookup and its event durations disagree. **Notes** Licences, from each repository’s record: Bak CC BY 4.0, MC_RTT CC BY 4.0, ds007420 CC0, FALCON H2 CC BY 4.0, Rabut et al. CC BY-NC 4.0 (non-commercial). Bak’s ethics and consent statement was read from an Internet Archive capture of the publisher’s article page, because the publisher refused automated requests. MC_RTT is animal data collected under procedures approved by institutional animal care and use committees, so informed consent does not apply. Not published: per-person, per-recording, per-channel, per-trial and per-row values; confusion matrices; signals, features, predictions and coefficients; measured compute; demographics; and the handoff, reviews and independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. **Review manifest SHA-256** `ae02c2115414c0ca61cf5c72a748a004f0f2854570570e56935515922de2070f` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [beyond-scalp-update.json](https://bci.report/data/beyond-scalp-update.json) | 38.8 KiB | `fe37a2a5c2105cdeadd6e3fb28b7944ca626fec4743e4c813d5aa669182b4341` | 8 October 2026 `questions-in-language-update-20261008` Route 3 of the decision-research roadmap, questions in language: can an EEG model be asked its questions in words, including ones it was never trained on? One small question-conditioned head over frozen features is asked each question by a question number, a label template or a natural-language description. SSVEP on BETA (a plain spectrum and frozen CBraMod features) and sleep on BOAS (frozen CBraMod features) are primary, three seeds each; EESM19, OpenBMI motor imagery, REVE-L features, a second text encoder, Chinese wordings, a negation probe, a contiguous held-out band of frequencies and Wearable-102 are secondary. Seen questions in their training wording, rewordings and questions the EEG head was never trained on are reported apart, never pooled into one zero-shot number, and every entry carries two flags: whether its interval excludes zero, and how it stands against a margin fixed before any result (2 pp for accuracy; a ratio of remaining error of 1.25 for the unseen sleep questions; the primary comparisons with the numeric code, the neighbour average and the shuffled templates are read for a difference only). “Jev-style” here means the interface of a decision model like TypeSafe’s Jev, applied to EEG: encode the recording once, then answer several explicit, typed questions about it, each with a probability. The roadmap started from a vision paper (Yu & Yao, 2026) that applies that idea to images. **Pages** [Questions in language](https://bci.report/topics/questions-in-language/) · [When not to act](https://bci.report/topics/when-not-to-act/) **Notes** Mostly a negative, boundary result. On sleep, asking a seen question with a template or a description was equivalent to a question number within 2 pp; on SSVEP it cost accuracy with both kinds of features. A new sentence frame cost nothing measurable on sleep, while synonyms and new descriptions lowered accuracy on both datasets (on SSVEP with frozen CBraMod features the synonyms stayed within the margin). Questions the EEG head was never trained on were not answered well: in every wording, each unseen sleep question did worse than adding up the head’s own stage answers, and every comparison of an unseen flicker frequency with CCA, which needs no training, fell far short. Negated wordings were answered as if they asked for what they negate. **Notes** Before any result, the pre-run permutation canary failed twice (revisions 3 and 4; every failing cell was on BOAS, one of them below chance). The owner overrode revision 4’s “no further revision” and approved revision 5, which passed; revision 4’s centring diagnostic was not run, and both failed records are kept and disclosed. For the Chinese text encoder the SSVEP token rule was reported instead of stopping the run (decision S0b-6); an engineering amendment before the first fit (EA-1) changed no question, partition, wording, arm, metric, margin or gate; the declared numeracy prediction was not activated, because its premise did not hold. There are 35 pre-declared primary comparisons, not corrected for multiplicity, and most secondary results are single-seed and were declared likely inconclusive before the run. **Notes** BOAS figures are published under the owner’s approval of 7 October 2026, extended to route 3 the same day (D11): with its three stated gaps and its attribution, participants pseudonymised in the public release, and every BOAS cell pooling at least 20 people; neither Bitbrain’s headband nor its automatic scoring is used or evaluated. Not published: per-person, per-night and per-fold values; the S10 per-frequency and interior-only breakdowns, which no independent audit covers; any value of a pre-run check, which is published as pass or fail and counts only; measured compute; and the handoff, the aggregates, the protocol and the four independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. The wording lists, the held-out partition and the derangements of the shuffled control are published in their own file. **Review manifest SHA-256** `c85dbadaab3d08b9ae079015e664cec4ed7e16438c6e1ca00689c1ef341ade5e` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [questions-in-language-update.json](https://bci.report/data/questions-in-language-update.json) | 409.9 KiB | `2583b5958c24e02898929c354e16653e50e33704aaaee37927e158f9d72d01dd` | | [questions-in-language-wordings.json](https://bci.report/data/questions-in-language-wordings.json) | 74.9 KiB | `fb8451a84f2d1d8a6a451e2669503656beb86f28edf9f02d73107ff5a9482529` | 8 October 2026 `later-sessions-update-20261008` Four results under one question — does a decoder trained on an earlier session still work later? — each from its own approved release, and none uses labels from the later session. WBCIC-SHU (62 people in two separate cohorts, 51 with two-class and 11 with three-class motor imagery): a source-majority prior and a relative spectral ridge trained on each person’s recording session 1 and tested on session 3, and frozen CBraMod with a ridge readout fitted on session 1, on the same people and trials. A longitudinal RSVP dataset (15 people): one ERP baseline trained at the first visit and scored at the publisher’s nominal Day 7, 80 and 200 visits. Forenzo’s continuous cursor-tracking dataset (23 admitted records in two cohorts, two response arms): a fixed spectral ridge from each record’s earliest to its latest complete session, against a source-mean comparator, published as a negative result. Nothing is pooled across datasets, cohorts or response arms, and the results share no ranking. **Pages** [Later sessions](https://bci.report/topics/later-sessions/) **Notes** WBCIC-SHU: in both cohorts the spectral ridge is above the source prior and frozen CBraMod above the spectral ridge, each with a paired interval above zero. The CBraMod arm was frozen after the spectral ridge results existed and reuses their test trials; there is no matched random-weight control, and whether this checkpoint saw WBCIC-SHU in pretraining is not established, so it does not show that pretraining caused the gain. Session numbers are recording-session ordinals, not a guaranteed time gap. **Notes** Longitudinal RSVP: AUROC is lower at the nominal Day 200 visit than at Day 7, with a paired interval below zero, and some people lost 0.05 AUROC or more; targets are rare, so average precision is printed beside AUROC. Day 7, 80 and 200 are the publisher’s nominal labels, and nothing here shows that elapsed time caused the change. Forenzo: the spectral ridge has a higher overall error than a constant source-mean comparator on every admitted record, in both cohorts and both response arms, with means far above medians; velocity imitates the publisher’s historical decoder output, not intended motion, and no online-control claim is made. **Notes** All three datasets are new to this site. Licences, from each repository’s record: WBCIC-SHU CC BY 4.0, the RSVP dataset CC0, Forenzo’s KiltHub record CC BY 4.0; each paper states ethics approval and written or signed informed consent. Not published: per-person and per-record values (minimums, percentiles and medians, except the Forenzo ridge medians the handoff requires beside their means), confusion matrices, the Forenzo recorded-decoder strata, measured compute, and the handoffs, release decisions and independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. **Review manifest SHA-256** `ece37ea382fbb7fc0fb2bfe360e47d673fd150862c909d7026b57c2e2af8a53e` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [later-sessions-update.json](https://bci.report/data/later-sessions-update.json) | 75.7 KiB | `b9db90486ef70f0c70b7feca315c58b09917e9327a1107d09b6f11b0d9982285` | 7 October 2026 `shared-representation-update-20261007` Route 2 of the decision-research roadmap, one representation and several questions: can one shared encoder with a fixed linear head per question answer several questions as well as a question-conditioned head, or as separate models, for less? Independent models, fixed heads, a shared hidden layer, and that layer conditioned on the question’s identity, compared at matched data and compute: EEGNet trained from scratch and frozen CBraMod features as the primary levels, three seeds each, and CBraMod adapted by LoRA, one seed, as a secondary level. Motor imagery on OpenBMI and sleep on BOAS are primary; EESM19 is a crude one-seed replication. Every contrast carries two flags — whether its interval excludes zero, and whether it lies within the 2 pp margin fixed before any result — and nothing is ranked where intervals overlap. **Pages** [One model, several questions](https://bci.report/topics/shared-encoder/) · [When not to act](https://bci.report/topics/when-not-to-act/) **Notes** The fixed-heads sentence is supported on neither domain: the question-conditioned head was higher on imagery against rest, and on five-stage sleep scoring, while the fixed heads were the cheaper set-up in both. The question’s identity itself added nothing measurable: the same hidden layer without conditioning was equivalent to the conditioned head within 2 pp on both motor-imagery questions and non-inferior on sleep. On sleep that layer carries the lift over the fixed heads; on motor imagery its own contrast with the fixed heads was inconclusive. Against a separate model per question, sharing one small EEGNet trunk cost accuracy on which hand; on imagery against rest and on the sleep stage that comparison was inconclusive. Questions at floor are reported and never counted, and with 21 primary entries and no multiplicity correction about one in twenty entries with no true difference may show one by chance. **Notes** BOAS is new to this site. Its rights review was approved by the owner on 7 October 2026 as publishable with stated gaps. The consent statement does not say whether participants agreed to public sharing or secondary use. The ethics and consent statements come from the publisher’s dataset description and README. No peer-reviewed paper describes BOAS. The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. Only cohort aggregates of at least 20 people are published. Participants are pseudonymised in the public release, and neither Bitbrain’s headband nor its automatic scoring is used or evaluated. **Notes** E2 sleep — CBraMod adapted by LoRA on BOAS — ran on 7 October 2026 by the owner’s decision, after every other result was known; its design and run condition were fixed at the freeze and nothing about it was chosen from results. It ran with temporary private resume checkpoints: none outlived its fit, and none remains. Not published: per-window and per-person values, including the per-person percentiles the release candidate carries for OpenBMI; measured compute totals (GPU-hours, wall time per fit and per block, training steps per outer fold), while the cost ledger publishes, labelled indicative, the two times measured interleaved on a shared GPU; and the aggregates and the three independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. **Review manifest SHA-256** `d777e980a42db43b14004677d3a93455414c62f45eaec51e73c7ab60bee103b4` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [shared-representation-update.json](https://bci.report/data/shared-representation-update.json) | 219.1 KiB | `0dec02f86d2ef9347c4e42e8c2c7d165ad169293247b759049b170f0f2ae41d2` | 4 October 2026 `foundation-models-update-20261004` The v9 foundation-model evaluation: eleven further EEG foundation models — sixteen encoder checkpoints — run as frozen probes on the eight core protocols with the published LaBraM and CBraMod recipe and only the encoder swapped (the same windows, people, folds, heads and scoring), and nine of them adapted on EEGMAT with the 1 October recipe, a head trained on the frozen encoder against rank-4 LoRA over three seeds. New rows beside the core matrix, in their own files: one JSON, and a CSV per protocol whose first columns are the core results CSVs’. Grouped by family, never ranked: a row is above or below another only where their 95% intervals do not overlap. Every cell says whether its dataset is in the model authors’ published pretraining list, checked on 4 October 2026 with the source, and every model carries its weights licence. **Pages** [Protocols](https://bci.report/protocols/) · [Methods](https://bci.report/methods/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) · [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) **Notes** Sleep staging is the only protocol where new frozen cells — ST-EEGFormer Large and Base, and REVE Large — lie entirely above every published row. On the other six scored protocols no new cell lies entirely above the published non-foundation row with the highest point estimate; on BETA, no new row lies above standard CCA, which needs no training. On EEGMAT, frozen readouts and LoRA order the models differently: one fixed recipe on one task, with LoRA budgets that differ by model, not a ranking. **Notes** BETA is in ST-EEGFormer’s published pretraining list and TMNRED in SingLEM’s, so those cells are flagged; ZUNA 1.1 does not list its pretraining data, so every ZUNA 1.1 cell is marked unknown. Every other cell, LaBraM’s and CBraMod’s included, is not in the authors’ published pretraining list as checked on 4 October 2026: a sourced statement, not proof that the recordings were never seen. Weights licences travel with the rows: REVE under its Responsible Use License, LUNA under CC BY-ND 4.0, ERP-FM for non-commercial use only, and ZUNA 1.1’s model card limits it to research use, not diagnosis or clinical use. **Notes** Not published: per-trial, per-person and per-fold values, features, and timings and memory — measured on a shared GPU, so the CSVs leave the scoring-time column empty — the declared sensitivity runs beyond the row footnotes that state them, and the aggregate, the group summaries and the six independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. Two BrainOmni cells were not run, because its tokenizer needs two-second windows: they are null with that reason, never zero. The released matrix keeps its bytes; the new rows never enter experiments.json. **Review manifest SHA-256** `234fc9bf732de4e5cd7b1c94456233d11f7621a5dbc42a615f84fe9b382bcf57` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [foundation-models-update.json](https://bci.report/data/foundation-models-update.json) | 325.3 KiB | `eaaaaccb2e8b70cacc85db71726d63de036c83c3d60b7b4a9097f9aafafad13a` | | [foundation-models-mi-rest.csv](https://bci.report/data/foundation-models-mi-rest.csv) | 19.0 KiB | `29f29270ca80637936b3a54fe7f1d8b000516329b9584b90265f4c0d5a481775` | | [foundation-models-idle.csv](https://bci.report/data/foundation-models-idle.csv) | 20.4 KiB | `9c32e21661d602c8c7cb90ee08617c1ce34dc69afe60e03731c3566236d37464` | | [foundation-models-beta-8ch.csv](https://bci.report/data/foundation-models-beta-8ch.csv) | 22.4 KiB | `8ed510e75ec78e6db8f81c5b3d3c9c53bc92188ac59d27936f69b6fca01d88f9` | | [foundation-models-beta-4ch.csv](https://bci.report/data/foundation-models-beta-4ch.csv) | 22.2 KiB | `56e716518fe5383bcabe50977a588082d8ef780486569740cb80f62392bbf9be` | | [foundation-models-arithmetic-rest.csv](https://bci.report/data/foundation-models-arithmetic-rest.csv) | 22.0 KiB | `61cd05ee7bf418e759663632942e57da8a0b4ded6b2156aa3885fe0c2005db7c` | | [foundation-models-p300-target.csv](https://bci.report/data/foundation-models-p300-target.csv) | 19.9 KiB | `45e9bb5e87ed07e94d783078416171a1da05b6686a5653517f004f570f6d96c4` | | [foundation-models-semantic-target.csv](https://bci.report/data/foundation-models-semantic-target.csv) | 22.2 KiB | `011e724ab480305b98cc8d910e0d83df7f65d94123bb8620ab0a88529c76b115` | | [foundation-models-sleep-scalp.csv](https://bci.report/data/foundation-models-sleep-scalp.csv) | 23.1 KiB | `9128234574c9c49152db7b8a1c052b65ad42f5cb56afab57dc97f4e9ab2b5e7d` | 4 October 2026 `reliable-decisions-update-20261004` Route 1 of the decision-research roadmap, reliable decisions: when a decoder should decline to decide, and what its confidence is worth. The saved test scores of models this site already publishes — spectral ridge, EEGNet and two LaBraM arms on EEGMAT (36 people); standard CCA, CBraMod and EEGNet on BETA (70 people) — rescored under a fixed confidence threshold, a coverage target, a certified selective risk and a learned reject option, with probability quality and a person-specific recalibration whose label cost is counted. No classifier was retrained, so every full-coverage accuracy equals the published one. Every protocol is balanced or uniform by design: these are method comparisons, never deployment error rates, and nothing is ranked where intervals overlap. **Pages** [When not to act](https://bci.report/topics/when-not-to-act/) **Notes** The certified selective risk certified no outer fold on EEGMAT, so it accepted nothing there: no error rate, not a zero one. On BETA its realised error exceeded a certified fold’s own target in some folds: a nominal guarantee across people, whose failure rate is measured here, not assumed. The learned reject option had a lower error than the model’s own calibrated confidence for no method. **Notes** ds003810 (ten people) is shown as crude, in a collapsed robustness panel beside secondary seeds, two sensitivity arms (their risk-certification fold counts only) and an exploratory reject head trained jointly with its classifier. Left out: the idle protocol, whose route-1 window proportions are a different unit from the trial-level idle figures on the same page, and every figure of the BNCI2015-001 arm, whose editorial hold stands. **Notes** Not published: per-trial, per-person and per-fold values, including the release candidate’s person-level percentiles and its median temperature over folds; the accuracies of the matched-holdout, map-sensitivity and jointly trained arms; and the independent audits themselves, pinned by hash in the review manifest because they carry private storage paths. The primary audit’s one failed check is a documentation error, not a computation error, and is stated as such. **Review manifest SHA-256** `49486a87b692e03ddaad411c140a533816fba3b34fd19ce261faa243d6cd2885` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [reliable-decisions-update.json](https://bci.report/data/reliable-decisions-update.json) | 114.1 KiB | `9af91aa919e6af66bb2907c913ce616b2f98337bf6c9f11bba490d75b7f9fe51` | 3 October 2026 `large-source-update-20261003` Two new sources, two separate questions, each with two fixed classical CPU baselines and no shared ranking. Dreem: five-stage sleep staging against the publisher’s consensus in 25 healthy sleepers (DOD-H) and 55 people with obstructive sleep apnoea (DOD-O), kept as separate experiments — a training prior and a spectral ridge, accuracy beside balanced accuracy, macro F1, Cohen’s kappa and the ridge’s stage-by-stage results. And OpenBMI: a motor-imagery decoder trained on a person’s first session and tested on their second after 0, 10, 20 or 40 labelled trials from it, in 51 people, with how many people declined beside every mean change. No foundation-model or fine-tuning result. **Pages** [Sleep-stage balance](https://bci.report/topics/sleep-staging/) · [How much calibration?](https://bci.report/topics/calibration-budget/) **Notes** The spectral ridge never predicts N1 in either cohort, so its N1 precision is not defined: it is published as null, never as zero. The training prior’s balanced accuracy sits slightly above one fifth because a night’s mean covers only the stages that night contains; it is a floor, not a chance level. Neural-network and foundation models on the Dreem cohorts are held, because the source’s physical units disagree. **Notes** OpenBMI is an expanded cohort under the same fixed method — the original 40 people with their results unchanged, plus 11 later-eligible people — not an independent replication; an earlier 40-person snapshot was prepared but never published here and is not a separate result. The relative-PSD change at 40 trials has an interval that includes zero and is not established. **Notes** Not published: per-person and per-night values (including the worst observed change, medians and 10th percentiles), Brier scores, log-likelihoods and calibration errors, the combined DOD-H plus DOD-O summary, and the independent audits themselves, which are pinned by hash in the review manifest because they carry private storage paths. **Review manifest SHA-256** `3126ca1b52f5711c7ae8227c7abab6e4e5cdd917944675cd00e414e62eb08c1e` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [large-source-update.json](https://bci.report/data/large-source-update.json) | 49.5 KiB | `834411fc728d2596d4e48be533f2409921e307e5ebbb248300c9fb4cf722cbcd` | 2 October 2026 `extension-update-20261002` Twenty further people of the asynchronous SSVEP release, scored with a rejection threshold fixed on the four-person pilot and with a personal one: detection, coverage, correct-and-accepted output and false acceptance per non-control state, side by side. And a new source, LTRSVP: a P300 decoder trained on one recording at 5 or 10 images a second and tested on a different, 10-Hz recording, with the full rate-by-rate matrix. Fixed classical baselines; no foundation-model or fine-tuning result. **Pages** [When not to act](https://bci.report/topics/when-not-to-act/) · [Screen to VR](https://bci.report/topics/screen-to-vr/) **Notes** The personal threshold raised detection balanced accuracy for 10 of 20 people and lowered it for 8, and did not raise the share of commands both accepted and correct. The image-rate difference has an interval that crosses zero, and rate and recording change together, as does time in the session (the original study presented the rates from the lowest to the highest, not randomised), so no rate effect is claimed. **Notes** Not published: per-person values, thresholds, predictions, and the independent audits themselves, which are pinned by hash in the review manifest because they carry private storage paths. **Review manifest SHA-256** `2dacb0131128304a9ece7bb8e8591819d36c85cd97ec7c5a3c5a0f077cfa2653` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [extension-update.json](https://bci.report/data/extension-update.json) | 23.4 KiB | `1c2eba3138b53f517df1ea467c30ce8bebbaee86873f92e1d1133b1278bb37a3` | 1 October 2026 `adaptation-update-20261001` New people, same task, zero labels from the test person: LaBraM adapted on mental arithmetic three ways — head only, last block, rank-4 LoRA — with the same checkpoint, folds, starting heads, batch order and five-epoch recipe, over three seeds. The next-day experiment is listed as status only. **Pages** [Which part to update?](https://bci.report/topics/model-adaptation/) **Notes** Printed beside the core matrix's frozen LaBraM readout on the same people and folds, because this batch's head-only arm is a short gradient-trained head and not the strongest frozen readout. **Notes** Not published: memory figures (allocator samples are lower bounds, not per-method peaks) and any figure from the next-day experiment. **Review manifest SHA-256** `5e53c5a4ef378a486ef94b6f7479e79c7310188c433933866e445a1a5b108068` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [adaptation-update.json](https://bci.report/data/adaptation-update.json) | 15.2 KiB | `6253e1d7a7915e12afdfe926688a0bbeb52f2fef28c21d8423c406e353902bc2` | 27 September 2026 `context-update-20260927` P300 calibration carried from a PC screen to a VR headset and back for 21 people; walking-speed scores from dry-electrode EEG printed beside a movement-nuisance comparator; and a four-person pilot of SSVEP windows accepted when no command was intended. Fixed CPU baselines; no foundation-model or fine-tuning result. **Pages** [Screen to VR](https://bci.report/topics/screen-to-vr/) · [On the move](https://bci.report/topics/on-the-move/) · [When not to act](https://bci.report/topics/when-not-to-act/) **Notes** Not published although correct: the direction-specific transfer means, because the independent audit did not recompute them. **Review manifest SHA-256** `376ac7a308d788c7a8d71a63ee7de5926906f07d3933f4f6a89da0818a70617d` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [context-update.json](https://bci.report/data/context-update.json) | 15.2 KiB | `4eb7eb4ef11ab0853ece1728cbf3708f1eb66be35ba8adbbaf840c579512fc07` | 23 September 2026 `clinical-update-20260923` A 149-person clinical case/control comparison, printed beside an age-and-sex-only comparator, and three holds that produced no score. **Pages** [Clinical groups](https://bci.report/topics/clinical-groups/) **Notes** Not published although aggregate: the per-group age and sex composition of the clinical cohort. **Review manifest SHA-256** `aab3a961abc4f0f82efec911cb28b5b2e0927040545ea41f42b897da98765572` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [clinical-update.json](https://bci.report/data/clinical-update.json) | 11.0 KiB | `082a25d15436eb52ebf1e25f981088f2fa56f21377f7ed0e7907d849a33678df` | 22 September 2026 `evidence-update-20260922` In-ear against scalp sleep staging, four posterior electrodes against sixteen, a physical-phantom motion test, and the status of planned adaptation experiments. **Pages** [Fewer electrodes](https://bci.report/topics/fewer-electrodes/) · [On the move](https://bci.report/topics/on-the-move/) · [Which part to update?](https://bci.report/topics/model-adaptation/) **Notes** Amended the EESM19 consent note (see data use). The core snapshot and the topic export were regenerated with one added citation; no score changed. Their current hashes, below, are the amended files. **Review manifest SHA-256** `acf92eb64069f0145130e500b75766b4e72a69c07fd68d77958340db828851a4` | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [evidence-update.json](https://bci.report/data/evidence-update.json) | 17.6 KiB | `219b6d4633894a83f12dd7e1e1ae8820ce9718ddca6807a57f49cf13337ba51a` | 20 September 2026 `deployment-topics-20260920-v1` Four deployment questions, each with its own protocol: paired contrasts, descriptive intervals and three-seed sensitivity groups. **Pages** [Dry vs. wet electrodes](https://bci.report/topics/dry-vs-wet/) · [On the move](https://bci.report/topics/on-the-move/) · [How much calibration?](https://bci.report/topics/calibration-budget/) · [Does pretraining help?](https://bci.report/topics/does-pretraining-help/) | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [deployment-topics.json](https://bci.report/data/deployment-topics.json) | 98.2 KiB | `708e6daad94bad3e32a4602af276e40d559e262741e92196a3590a74c6d9232f` | 20 September 2026 `research-preview-20260920` The core matrix: eight fixed protocols on seven public data sets, with one file of results and one protocol descriptor per protocol. **Pages** [Core matrix (home page)](https://bci.report/) | File | Size | SHA-256 of the file as served | | --- | --- | --- | | [experiments.json](https://bci.report/data/experiments.json) | 89.7 KiB | `113de335c17043bd9f720cf6e46663ff8bb2a11ddedaa49282765ed732807cd0` | | [mi-rest-results.csv](https://bci.report/data/mi-rest-results.csv) | 3.6 KiB | `50f0eef66b1d36314c5ec4a01028a30c58e2ff88ebac08325ee0b6c4158a19cf` | | [mi-rest-protocol.json](https://bci.report/data/mi-rest-protocol.json) | 4.2 KiB | `1c8af89bad4fe60b5afa21635802651395311b5dbca02d1f33717d6cef4fbc65` | | [idle-results.csv](https://bci.report/data/idle-results.csv) | 4.6 KiB | `d511d329755caf81109305a78e416a2650e84ee03f0fa965c34d3e8220796d81` | | [idle-protocol.json](https://bci.report/data/idle-protocol.json) | 2.8 KiB | `f648b56dfb1d030170440884ac66106ef369c7df8d27040c3bd8d985fc1f0ee8` | | [beta-8ch-results.csv](https://bci.report/data/beta-8ch-results.csv) | 4.6 KiB | `01a11e0ba22ea726158c936ff66a7e60f8704641a6e8f11d5c83f25d87fd3125` | | [beta-8ch-protocol.json](https://bci.report/data/beta-8ch-protocol.json) | 3.3 KiB | `ac384c95a33d068d38c37cafae7788331f37f5be5a296bfa01d056daf5590426` | | [beta-4ch-results.csv](https://bci.report/data/beta-4ch-results.csv) | 4.5 KiB | `9ffe23197b77eda5f08dca476f66197c2da77f198ce8dd35dfd0df972a24082d` | | [beta-4ch-protocol.json](https://bci.report/data/beta-4ch-protocol.json) | 3.3 KiB | `749d13951729e00935724fce9b790e8f23978e3290a9dd9df6e5a276736e69a1` | | [arithmetic-rest-results.csv](https://bci.report/data/arithmetic-rest-results.csv) | 3.7 KiB | `bed02323936f2980eac7bf3a8e8b2595630fa1609c1466600d33acc2f83928dc` | | [arithmetic-rest-protocol.json](https://bci.report/data/arithmetic-rest-protocol.json) | 4.4 KiB | `a2d8a9f182916b1ae611b163d65b3b7184bdbf79e4fa39ee47e778c02c301c6b` | | [p300-target-results.csv](https://bci.report/data/p300-target-results.csv) | 3.8 KiB | `6cd242313b592d37be2c1e6af2fd8524f612010547a143c84948a6f03a7a852d` | | [p300-target-protocol.json](https://bci.report/data/p300-target-protocol.json) | 3.0 KiB | `ba94ddcd7e8fb925d97ebfd7c0418b69e02d75a1b6135a9be0e058b0a7dfaee9` | | [semantic-target-results.csv](https://bci.report/data/semantic-target-results.csv) | 4.4 KiB | `b4baad3795b5512a033609512c90055894caf915b62c9c303fd77310d98caaaa` | | [semantic-target-protocol.json](https://bci.report/data/semantic-target-protocol.json) | 3.2 KiB | `b6c5a380596148ecfdd5bc98b2799c12dd5ceec0623c9ebdb8a711ee01fb7653` | | [sleep-scalp-results.csv](https://bci.report/data/sleep-scalp-results.csv) | 3.9 KiB | `55be9dc880b0c620f9dfba3f97c885e50645db0984de2b7d3c10a380fc9db32d` | | [sleep-scalp-protocol.json](https://bci.report/data/sleep-scalp-protocol.json) | 4.3 KiB | `6cf58be78f646b20f2ff9a7a9c0beadee4bfd24b9046ac68dc5f68afb2282036` | Each batch’s source terms, consent and privacy review: [Data use & privacy](https://bci.report/data-use/) --- Markdown copy of https://bci.report/releases/archive/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/ # Evidence you can inspect. Recordings stay with their sources. SOURCES & RESPONSIBLE USE BCI Report is a personal, noncommercial research project. It has no advertising, paid access, EEG upload form or diagnostic service. **Downloads contain aggregate benchmark results and methods.** They do not contain raw EEG, individual scores, participant identifiers, recordings, embeddings or model weights. Access original data through the credited source and follow that source’s terms. ## What a source review means A source review checks the original release: its stated license, attribution requirements and available consent or ethics documentation. A Hugging Face mirror is a distribution source, and its metadata alone does not establish the uploader’s rights or participants’ consent. Numerical audits and permission to publish are separate checks. “Aggregate results” means this project selected cohort-level outputs for the stated research scope. It implies no legal certification, permission to reuse raw data, verified anonymity of the source files or endorsement by the original researchers. ## Core and deployment sources The deployment extension carries nine reviewed citation records: six for datasets already in the matrix or research register, and three for the wearable and mobile sources. | Dataset | Source terms | Scope | | --- | --- | --- | | Stable 39-comparison matrix and research register | | | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate results** — Clear CC0 snapshot plus dataset-specific ethics and signed-consent evidence; aggregate-only publication materially limits privacy exposure. | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate results** — Attribution license and study-specific approval/consent are documented; eligibility assumes aggregate-only output and exclusion of subject-info fields. | | [EESM19 scalp subset ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. | [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate results** — Clear upstream CC0 and study ethics evidence; use only the pinned scalp-signal derivative and publish aggregates, with provenance disclosed. Consent covered the study, not publication; the public release rests on a GDPR anonymization assessment (amended 2026-09-22). | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate results** — Eligible only under the stated personal, research-led, noncommercial operation. Any ads, sponsorship, fees, or use directed toward commercial advantage requires fresh review or permission. | | [BNCI2014-008 ↗](https://bnci-horizon-2020.eu/database/data-sets/008-2014) — BNCI Horizon 2020 catalog · Original researchers credited at source. | [CC BY-NC-ND 4.0 ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/) | **Not in this release** — Small clinical cohort and linked health attributes create disproportionate reidentification risk; aggregate-only output does not resolve the missing secondary-publication consent evidence. | | [BNCI2014-009 ↗](https://bnci-horizon-2020.eu/database/data-sets/009-2014) — BNCI Horizon 2020 catalog · Original researchers credited at source. | [CC BY-NC-ND 4.0 ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/) | **Not in this release** — License permits only noncommercial unadapted redistribution, and the consent/ethics chain for public secondary results remains unverified. | | [stew-monster-processed ↗](https://ieee-dataport.org/open-access/simultaneous-task-eeg-workload-dataset) — monster-monash/STEW Hugging Face processed mirror · Original researchers credited at source. | [Mirror declares CC BY 4.0; exact upstream DataPort license not independently verified ↗](https://creativecommons.org/licenses/by/4.0/) | **Not in this release** — Ethics evidence is positive, but upstream license and the processed mirror's authority to apply CC BY remain unresolved. | | [TMNRED / ds005383 ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) — Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0. | [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate results** — Strong study-specific open-sharing evidence. Aggregate metrics avoid redistribution of stimulus text and participant metadata; credit under CC BY. | | [ds006593 ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) — OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate results** — Pinned CC0 release with dataset-specific IRB and consent evidence and comparatively low metadata exposure. | | [physionet-eegmmidb-1.0.0 ↗](https://physionet.org/content/eegmmidb/1.0.0/) — PhysioNet EEG Motor Movement/Imagery Dataset 1.0.0 · Original researchers credited at source. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Not in this release** — Reuse license is clear, but the dataset-specific human-subject consent and ethics chain is not evidenced strongly enough for a new public benchmark release. | | [BNCI2014-001 ↗](https://bnci-horizon-2020.eu/database/data-sets/001-2014) — BNCI Horizon 2020 catalog · Original researchers credited at source. | [CC BY-ND 4.0 ↗](https://creativecommons.org/licenses/by-nd/4.0/) | **Not in this release** — Per-dataset license is verified, but consent evidence and the application of ND to the planned publication pipeline should be clarified. | | [BNCI2014-004 ↗](https://bnci-horizon-2020.eu/database/data-sets/004-2014) — BNCI Horizon 2020 catalog · Original researchers credited at source. | [CC BY-ND 4.0 ↗](https://creativecommons.org/licenses/by-nd/4.0/) | **Not in this release** — Per-dataset license is verified, but consent evidence and the application of ND to the planned publication pipeline should be clarified. | | [ds005342 ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) — OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate results** — Eligible only as an aggregate-only case study: remove all participant IDs/results and condition breakdowns, label n=4 prominently, avoid subgroup claims, and make no claim that the cohort is anonymous. | | Deployment-topic extension | | | | [An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) — Official Tsinghua BCI Lab author mirror snapshot acquired 2026-09-20; the mirror itself has no separate DOI or version number. | CC BY 4.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [Lee2021 – SSVEP paradigm of the Mobile BCI dataset ↗](https://nemar.org/dataset/nm000125) — NEMAR/MOABB-generated BIDS derivative | CC BY 4.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [ERP paradigm of the Mobile BCI dataset ↗](https://nemar.org/dataset/nm000201) — NEMAR/MOABB-generated BIDS derivative with an embedded sourcedata provenance record | CC BY 4.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — Processed mirror identifies the canonical Figshare record and repeats CC BY 4.0; upstream version is pinned separately. | CC BY 4.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Direct OpenNeuro snapshot; no mirror in the reviewed acquisition path. | CC0-1.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [TMNRED / ds005383 ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) — Direct OpenNeuro snapshot; publication also points to author-hosted raw/preprocessed data. The license-label mismatch is conservatively resolved in favor of attribution. | OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [ds006593 ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) — Direct OpenNeuro snapshot; NEMAR independently identifies its copy as derived from this OpenNeuro record. | CC0-1.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [EESM19 scalp subset ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Processed H5 mirror derived from OpenNeuro ds005185. The mirror card agrees on CC0, but the derived-file transformation is described by the mirror rather than independently certified by upstream. | CC0-1.0 declared by upstream and mirror | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Direct pinned PhysioNet release. | Open Data Commons Attribution License 1.0 | **Aggregate topic results** — Participant-level scores, demographics and recordings stay outside the site. | ## Sources added 22 September 2026 Each sits outside the eight-protocol matrix, with its own recorded editorial decision and aggregate export. This and every later batch on this page has its own review manifest, whose SHA-256 the releases page lists. | Dataset | Source terms | Consent and privacy | | --- | --- | --- | | [EESM23 · Ear-EEG Sleep Monitoring 2023 ↗](https://doi.org/10.18112/openneuro.ds005178.v1.0.0) — Dataset: Yousef Rezaei Tabar, Kaare Mikkelsen, Laura Birch, Nelly Shenton, Simon L. Kappel, Astrid R. Bertelsen, Reza Nikbakht, Hans O. Toft, Chris H. Henriksen, Martin C. Hemmsen, Mike L. Rank, Marit Otto and Preben Kidmose · Ear-EEG Sleep Monitoring 2023 (EESM23), OpenNeuro ds005178 v1.0.0, doi:10.18112/openneuro.ds005178.v1.0.0. Study: Kaare Bjarke Mikkelsen, Yousef Rezai Tabar, Laura Rævsbæk Birch, Simon Lind Kappel, Christian Bech Christensen, Lars Dalskov Mosgaard, Marit Otto, Martin Christian Hemmsen, Mike Lind Rank and Preben Kidmose · Ear-EEG sleep monitoring data sets, Scientific Data 12, 301 (2025), doi:10.1038/s41597-025-04579-8. Processed export: Zachary1150/EESM23-Processed on Hugging Face, which did not create the original cohort. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | — All subjects gave written, informed consent. Approved by the Central Denmark Region Committees on Biomedical Research Ethics (1-10-72-13-20) and the Danish Medicines Agency (2020012619). Publication was not mentioned in the informed consent form; before release, the GDPR office of Region Midt judged the data set fully anonymized. Consent therefore covered the study, and the public release rests on that anonymization judgment. Only cohort aggregates are published here: no per-person scores, no per-person distribution percentiles, no per-night or per-epoch values. | | [Phantom EEG with motion, muscle and eye artifacts ↗](https://doi.org/10.18112/openneuro.ds004784.v1.0.4) — Ryan J. Downey and Daniel P. Ferris · Phantom EEG Dataset with Motion, Muscle, and Eye Artifacts and Example Scripts, OpenNeuro ds004784 v1.0.4, doi:10.18112/openneuro.ds004784.v1.0.4. Method: doi:10.3390/s23198214. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | — One electrically conductive physical head phantom. No human participants are represented, so consent and ethics review do not apply; no population inference is made. | | [EEG Alpha Waves dataset ↗](https://doi.org/10.5281/zenodo.2605110) — Grégoire Cattan, Pedro L. C. Rodrigues and Marco Congedo · EEG Alpha Waves dataset, Zenodo, doi:10.5281/zenodo.2605110. Report: Grégoire Cattan, Pedro Luiz Coelho Rodrigues and Marco Congedo · EEG Alpha Waves Dataset, research report, GIPSA-lab, Grenoble, 2018, hal-02086581. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | — The primary report states that all participants provided written informed consent confirming the notification of the experimental process, the data management procedures and the right to withdraw from the experiment at any moment. It names no ethics committee or approval number, and it does not say whether public release was described in the consent; the authors released the recordings openly under CC BY 4.0. The source also ships a per-person age, gender and fatigue table; this project did not download it. Only cohort aggregates are published here: no per-person scores, no per-person distribution percentiles, no per-recording values. | **Amended 22 September 2026 — EESM19.** Its 2025 data descriptor, which covers EESM23 as well, states that publication was not mentioned in the informed consent form. The public release rests on the GDPR office of Region Midt judging the data fully anonymized. No score, cohort or decision changed. ## Sources added 23 September 2026 Outside the eight-protocol matrix: one published as aggregate scores, two as status only. | Dataset | Source terms | What is published | | --- | --- | --- | | [OpenNeuro ds004584 · Rest eyes open, Parkinson's disease and controls ↗](https://doi.org/10.18112/openneuro.ds004584.v1.0.0) — Dataset: Arun Singh, Rachel Cole, Arturo Espinoza, Jim Cavanagh and Nandakumar Narayanan · Rest eyes open, OpenNeuro ds004584 v1.0.0, doi:10.18112/openneuro.ds004584.v1.0.0. Cohort study: Arun Singh, Rachel C. Cole, Arturo I. Espinoza, Jan R. Wessel, James F. Cavanagh and Nandakumar S. Narayanan · Evoked mid-frontal activity predicts cognitive dysfunction in Parkinson's disease, Journal of Neurology, Neurosurgery & Psychiatry 94, 945-953 (2023), doi:10.1136/jnnp-2022-330154. Transport mirror: NEMAR on004584, doi:10.82901/nemar.on004584. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Every participant gave written informed consent, and decisional capacity was established, in accordance with the Declaration of Helsinki; all research protocols were approved by the University of Iowa Human Subjects Review Board (IRB 201707828). One site, the University of Iowa Movement Disorders Clinic. The recordings are released CC0 by the authors. This is a clinical population, so the publication boundary is tighter than elsewhere on this site: only two model-level scores with descriptive intervals are published. No participant rows, per-person predictions, clinical or cognitive scores, medication state, or per-group age and sex composition — the last of these is in the reviewed inputs and is deliberately left out, because this site does not publish demographic profiles of a diagnosis-specific group. Nothing published here is diagnostic, and no participant can be distinguished in it. | | [OpenNeuro ds004902 · Paired sleep-condition recordings ↗](https://openneuro.org/datasets/ds004902/versions/1.0.8) — OpenNeuro ds004902 v1.0.8, cited as the pinned release while this hold stands. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Status only — no scores** — Downloaded in full and identity-verified, but five FDT payloads in the canonical release are shorter than their own headers and sidecars declare, and all five end inside an incomplete channel frame. The frozen 71-person paired protocol cannot pass its intake gate, so there is no score of any kind. Padding, truncating to the shortest common length, or reporting a partial cohort under the frozen protocol were all refused. What is published is the status and the source-defect counts; no accuracy row exists to publish. | | [L-FAME · six-person expansion ↗](https://huggingface.co/datasets/L-FAME-Dataset-Benchmark/L-FAME/tree/53f4a329eed86a0053b2150cc4e754406662b5a3) — L-FAME Dataset Benchmark · L-FAME, Hugging Face, at revision 53f4a329eed86a0053b2150cc4e754406662b5a3. Noncommercial scope; attribution required; the source card forbids re-identification attempts. | [CC-BY-NC-4.0 ↗](https://creativecommons.org/licenses/by-nc/4.0/) | **Status only — no scores** — An audited relative-spectrum description exists for 60 recordings from 6 people, and it is not a model result. Its per-condition numbers are withheld: the descriptors are summarised over 12 recordings per condition from those 6 people, so a minimum or maximum is one person's recording, which this site does not publish. Electrode-row mapping and physical units also remain unresolved, so a spectral profile cannot be read as a biological finding. The published record is the status and the cohort size. | **A clinical cohort is held to a higher bar.** Research results on a public data set. Not a diagnosis, not diagnostic accuracy, and not a medical device or medical advice. Not published, although both are aggregates in the reviewed inputs: the per-group age means and sex counts of a diagnosis-specific cohort, which is the demographic profile this policy excludes, and the per-condition band descriptors of the six-person L-FAME expansion, where a range bound is one person’s recording. The export code refuses all of them by field name. What each batch could not publish is in the [holds register](https://bci.report/releases/#holds), with the date each hold was opened and, where it was, resolved. ## Sources added 27 September 2026 Three published as aggregate scores, one as status only. The asynchronous SSVEP pilot was held until its data paper’s consent and ethics statement had been read, and released the same day. All four are fixed CPU baselines. | Dataset | Source terms | What is published | | --- | --- | --- | | [Cattan PC/VR P300 · a BCI experiment in virtual reality and on a personal computer ↗](https://doi.org/10.5281/zenodo.2605205) — Grégoire Cattan, Anton Andreev, Pedro L. C. Rodrigues and Marco Congedo · Dataset of an EEG-based BCI experiment in Virtual Reality and on a Personal Computer, Zenodo, doi:10.5281/zenodo.2605205; documentation arXiv:1903.11297. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The documentation reports that all 21 participants gave written informed consent covering the experimental process, the data management procedures and the right to withdraw at any moment, and that the study was approved by the Ethical Committee of the University of Grenoble Alpes (Comité d'Ethique pour la Recherche Non-Interventionnelle). Only cohort aggregates are published: means and participant-bootstrap intervals. The per-person minimum, maximum, median and quartiles in the reviewed inputs are dropped, because with 21 people each of them is one person's score. | | [Multimodal gait · EEG during treadmill walking at three speeds ↗](https://doi.org/10.13026/r0ea-7161) — R. Katmah, A. AlShehhi, D. Kosaji, N. Al-Rahmani, M. Abdullah, A. A. V. Hulleck and K. Khalaf · A multimodal gait dataset of brain activity, muscle activity, kinematics and ground forces in young adults, PhysioNet v1.0.0 (2026), doi:10.13026/r0ea-7161. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The PhysioNet record states that the study followed the Declaration of Helsinki, was approved by the Institutional Review Board of Khalifa University (protocol H19-038), and that all participants gave written informed consent. Only cohort aggregates are published: two model-level scores with participant-bootstrap intervals. No per-person or per-recording value, and no participant metadata. | | [YSU asynchronous SSVEP-BCI dataset ↗](https://doi.org/10.6084/m9.figshare.24906300.v3) — Jing Zhao, Qian Zhang, Xinrui Wang, Xueshuo Liu, Jiaxin Li, Fengjie Fan, Zhenhu Liang and Xiaoli Li · An EEG dataset for studying asynchronous steady-state visual evoked potential (SSVEP) based brain computer interfaces, Brain-Apparatus Communication 3(1) (2024), doi:10.1080/27706710.2024.2418650; data at doi:10.6084/m9.figshare.24906300.v3. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The data descriptor (section 2.1, Subjects) reports that before the experiment all 24 subjects thoroughly understood the experimental procedures and signed written informed consent, and that the study was reviewed and approved by the ethics committee of Qinhuangdao First Hospital. This project uses a four-person pilot of the release (the first four participant archives). Only pooled window counts are published. The per-participant ranges in the reviewed inputs are dropped: with four people, each bound is one person's rate. | | [Stieger longitudinal BCI · one-person pilot ↗](https://doi.org/10.6084/m9.figshare.13123148.v1) — James R. Stieger, Stephen A. Engel and Bin He · Continuous sensorimotor rhythm based brain computer interface learning in a large population, Scientific Data 8, 98 (2021); data at doi:10.6084/m9.figshare.13123148.v1. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Status only — no scores** — One person's eleven sessions were acquired, not the 62-person release. With one person, every score is that person's, which this site does not publish; the execution handoff also asks for no ranked score and no session table. What is published is that the loader and a chronological split work on this source, and how many trials meet the source's own eligibility rules. | ## Reviewed 1 October 2026 No new dataset: the LaBraM adaptation results reuse the 2026-09-20 review of the EEGMAT recordings unchanged, and the LaBraM Base checkpoint’s MIT-licensed repository was reviewed for aggregate evaluation on the same date. The checkpoint, adapters and trained heads are not redistributed. The second experiment ran on a source whose editorial hold stands, so it is listed as status only. | Dataset | Source terms | What is published | | --- | --- | --- | | [EEGMAT · LaBraM adaptation, three matched update rules ↗](https://physionet.org/content/eegmat/1.0.0/) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate scores** — Paper reports Bioethics Commission approval (15 August 2018) and written consent from every participant. EDF dates are normalized, but subject-info.csv contains age, gender, occupation, and recording date; those fields must not be published. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: cohort means, descriptive person-bootstrap intervals, per-seed cohort means, and counts of people helped, harmed or unchanged out of 36. No per-person or per-fold value. | | [BNCI2015-001 · same person, the next day ↗](https://bnci-horizon-2020.eu/database/data-sets) — Institute for Knowledge Discovery, Graz University of Technology · BNCI Horizon 2020 data set 001-2015, Autocalibration and recurrent adaptation: Towards a plug and play online ERD-BCI. | [CC BY-NC-ND 4.0 (as displayed by the official catalogue) ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/) | **Status only — no scores** — Run and independently replayed on 2026-09-22. The source's 2026-09-20 editorial hold stands: the official catalogue applies CC BY-NC-ND 4.0, and the dataset description gives written informed consent but no ethics-board approval. Status only until that review is complete; the hold is not a finding that aggregate statistics are prohibited. | Two pieces of released wording were also corrected on this date, a method description and a dataset credit, without changing any released file; both are listed on the [releases page](https://bci.report/releases/#corrections). ## Reviewed 2 October 2026 Both published as aggregate scores. The asynchronous SSVEP extension scores the remaining twenty people of the same figshare release and reuses by reference the consent and ethics record read from its data paper for the 27 September pilot. LTRSVP is a new source: the PhysioNet record carries no consent statement, so it was read from the original publication, which reports approval by the Ethics Committee of the University of Essex and written consent from every participant. Both are fixed classical baselines. Not published: the independent audits behind both, which carry private storage paths and are pinned by hash in the review manifest instead. | Dataset | Source terms | What is published | | --- | --- | --- | | [YSU asynchronous SSVEP-BCI dataset ↗](https://doi.org/10.6084/m9.figshare.24906300.v3) — Jing Zhao, Qian Zhang, Xinrui Wang, Xueshuo Liu, Jiaxin Li, Fengjie Fan, Zhenhu Liang and Xiaoli Li · An EEG dataset for studying asynchronous steady-state visual evoked potential (SSVEP) based brain computer interfaces, Brain-Apparatus Communication 3(1) (2024), doi:10.1080/27706710.2024.2418650; data at doi:10.6084/m9.figshare.24906300.v3. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Same release, same consent record: the data descriptor (Brain-Apparatus Communication 3(1), 2024, section 2.1, Subjects), read by the maintainer for the 2026-09-27 release, reports that all 24 subjects understood the procedures and signed written informed consent, and that the study was reviewed and approved by the ethics committee of Qinhuangdao First Hospital. Reused by reference from the 2026-09-27 manifest, not re-read. Published here: cohort means with participant-bootstrap intervals, pooled window counts, and the number of people helped, harmed or tied out of twenty. The per-participant ranges in the aggregate are dropped: each bound is one person's rate. | | [LTRSVP · EEG Signals from an RSVP Task ↗](https://physionet.org/content/ltrsvp/1.0.0/) — Ana Matran (University of Essex) · EEG Signals from an RSVP Task 1.0.0, PhysioNet (2017), doi:10.13026/C2KX0P. Original publication: Ana Matran-Fernandez and Riccardo Poli, Towards the automated localisation of targets in rapid image-sifting by collaborative brain-computer interfaces, PLoS ONE 12(5): e0178498 (2017), doi:10.1371/journal.pone.0178498. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate scores** — The original publication (Materials and methods, Participants and setup) states that the study received approval from the Ethics Committee of the University of Essex and that consent was obtained from all participants in written form before the experiment began; its data availability statement names this PhysioNet record. The PhysioNet record itself carries no consent statement. Published here: participant-mean balanced accuracy and AUROC for each of the nine training-rate and test-rate cells with participant-bootstrap intervals, one paired difference, and how many of the nine people were lower, higher or tied. No per-person, per-recording or per-image value, and none of the ages, sexes or handedness the paper reports. | ## Reviewed 3 October 2026 Both published as aggregate scores, each with two fixed classical baselines. The Dreem Open Datasets are a new source. The licence recorded is MIT, as declared in the Zenodo deposit’s archived publisher metadata. The upstream code licence could not be fetched, and nothing upstream is redistributed. The dataset paper gives its two cohorts different statements, and each is published as it stands (owner decision, 3 October 2026): for DOD-H the paper names an ethics approval and has no informed-consent sentence; for DOD-O it states written informed consent and names no ethics committee — the same basis on which Alpha Waves was published on 22 September. The physical units of the Dreem recordings are unresolved (the stored metadata says millivolts, the publisher’s converter microvolts), so models that need absolute amplitude are held. OpenBMI is the second new source: CC0 1.0 per the DataCite record GigaDB registered and GigaDB’s repository policy (the GPL-3.0 seen in third-party catalogues is the licence of the OpenBMI code, not of the data), and the paper states approval by the Korea University Institutional Review Board and written informed consent. Not published: the independent audits behind both, which carry private storage paths and are pinned by hash in the review manifest instead. | Dataset | Source terms | What is published | | --- | --- | --- | | [Dreem Open Datasets (DOD-H and DOD-O) ↗](https://zenodo.org/records/15900394) — Antoine Guillot, Fabien Sauvet, Emmanuel H. During and Valentin Thorey · Dreem Open Datasets: Multi-Scored Sleep Datasets to Compare Human and Automated Sleep Staging, IEEE Transactions on Neural Systems and Rehabilitation Engineering (2020), doi:10.1109/TNSRE.2020.3011181; preprint arXiv:1911.03221. Data: Dreem Open Datasets, Zenodo, doi:10.5281/zenodo.15900394. Official repository: github.com/Dreem-Organization/dreem-learning-open, revision 8b332d6827f5ae6a22f4bb97b4deef4273238ec3. | [MIT ↗](https://opensource.org/licenses/MIT) | **Aggregate scores** — The two cohorts carry different statements, and each is published as it stands. DOD-H, 25 healthy volunteers recorded at the French Armed Forces Biomedical Research Institute (IRBA): the paper states that the study was approved by the Committees of Protection of Persons (CPP), declared to the French National Agency for Medicines and Health Products Safety, and carried out in compliance with the French Data Protection Act, ICH standards and the Declaration of Helsinki; its DOD-H description has no informed-consent sentence. DOD-O, 55 people with obstructive sleep apnoea recorded at the Stanford Sleep Medicine Center (clinical trial NCT03657329): the paper states that all trial participants gave informed written consent before taking part; no ethics committee or review board is named for DOD-O. The authors present both as publicly available datasets, and the deposit declares open access. Published here: cohort-level participant-equal means with whole-night bootstrap intervals, cohort-level stage support counts, and paired differences. No per-night value, record name, fold assignment or demographic. | | [OpenBMI motor imagery (Lee et al. 2019) ↗](https://doi.org/10.5524/100542) — Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — The paper's Ethical Approval section states that the study was approved by the Korea University Institutional Review Board (1040548-KUIRB-16-159-A-2) and that written informed consent was obtained from all participants before the experiments. Published here: participant-equal means with whole-participant bootstrap intervals for each arm and calibration budget, paired contrasts, and how many of the 51 people declined. The release's per-person summaries (worst observed change, medians, 10th percentiles) are dropped, and no participant, session file or trial identifier is published. | **Dreem, cohort by cohort, as the paper states it.** DOD-H — ethics approval: Approved by the Committees of Protection of Persons (CPP); declared to the French National Agency for Medicines and Health Products Safety; carried out in compliance with the French Data Protection Act, International Conference on Harmonization (ICH) standards and the Declaration of Helsinki (1964, as revised in 2013). Informed consent: not stated. The paper's DOD-H description contains no informed-consent sentence. DOD-O — ethics approval: not stated. No ethics committee or institutional review board is named for DOD-O in the paper. Informed consent: All trial participants gave informed written consent before taking part. Read from arXiv:1911.03221v4 (27 April 2020), II. Materials and Methods, A. Datasets, read 2026-10-03. ## Reviewed 4 October 2026 Two reviews: route 1 of the decision-research roadmap, and the foundation-model evaluation. ### Route 1, reliable decisions Route 1 adds no new dataset and no new model: it rescored the saved test scores of models this site already publishes on EEGMAT, BETA and ds003810, whose rights records from the 2026-09-20 review are reused unchanged, and every full-coverage accuracy equals the published one. Published: pooled coverage and selective error, probability quality and paired contrasts with whole-person bootstrap intervals, and counts of people and of outer folds. Not published: per-trial, per-person and per-fold values, including the person-level percentiles and the median temperature over folds the release candidate carries; the route’s idle-protocol figures, which are in another unit; every figure of its BNCI2015-001 arm, whose editorial hold stands; and the audits behind the release, which carry private storage paths and are pinned by hash in the review manifest instead. ds003810 is shown as crude. | Dataset | Source terms | What is published | | --- | --- | --- | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate scores** — Paper reports Bioethics Commission approval (15 August 2018) and written consent from every participant. EDF dates are normalized, but subject-info.csv contains age, gender, occupation, and recording date; those fields must not be published. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: pooled coverage and selective error, probability-quality levels and paired contrasts with whole-person bootstrap intervals, and counts of people (with zero accepted windows, with a degenerate personal temperature) out of 36. No per-person, per-fold or percentile value. | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Paper reports Tsinghua University ethics approval 20190002 and written consent; parents consented for participants under 16. Ages span 9-64. Only cohort aggregates should be shown. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: pooled coverage and selective error, probability-quality levels and paired contrasts with whole-person bootstrap intervals, counts of people out of 70 and counts of outer folds out of 7. No per-person, per-fold or percentile value. | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Study paper reports ethics approval by CEySTE CCT-CONICET Santa Fe, Declaration of Helsinki compliance, and signed consent. Twelve adults participated; ten were retained. Public participants.tsv includes exact age, sex, handedness, and pseudonymous ID, which must not be surfaced. Reused unchanged from the 2026-09-20 review of the same recordings. Published here, in a panel labelled crude: the published balanced accuracy and the error when everything is accepted; under the fixed threshold, pooled coverage and the selective error (a count of wrong among accepted, flagged unstable, where fewer than ten were accepted), each with its whole-person bootstrap interval where defined, and how many of the ten people had nothing accepted; the learned-reject contrasts with whole-person bootstrap intervals; and the certified-risk rule's settings (delta and target) and fold counts, with nothing accepted: a coverage of zero, all ten people with nothing accepted, and so no error and no error relative to the target. Two people per test fold, so no per-fold block and no person-level spread. | ### Foundation models No new dataset, but new models. Eleven further EEG foundation models — sixteen encoder checkpoints — were run as frozen probes on the eight core protocols with the published recipe and only the encoder swapped, and nine of them were adapted on EEGMAT. The seven datasets behind those protocols keep their rights records from the 2026-09-20 review, reused unchanged. The new models’ weights carry their own terms, reviewed on the same date and listed below with the owner’s decision for each: the REVE Responsible Use License, accepted by the owner; LUNA’s CC BY-ND 4.0, approved for internal probing and adaptation with aggregate scores only and no endorsement implied; ERP-FM’s CC BY-NC-SA 4.0, non-commercial; and ZUNA 1.1’s model card, research use only, not for diagnosis or clinical use. No weights, adapted weights or LoRA deltas were kept or shared, and no model’s authors endorse these results. Pretraining exposure is what each model’s authors publish as their pretraining list, checked on 4 October 2026 and linked on the protocol and method pages: not proof that a recording was never seen. Not published: per-trial, per-person and per-fold values; features; timings and memory, which were measured on a shared GPU; and the aggregate, the group summaries and the six independent audits, which carry private storage paths and are pinned by hash in the review manifest instead. Release [`foundation-models-update-20261004`](https://bci.report/releases/#foundation-models-update-20261004). | Dataset · protocol | Source terms | What is published | | --- | --- | --- | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Motor imagery & rest · [mi-rest](https://bci.report/protocols/mi-rest/#foundation-v9) — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Study paper reports ethics approval by CEySTE CCT-CONICET Santa Fe, Declaration of Helsinki compliance, and signed consent. Twelve adults participated; ten were retained. Public participants.tsv includes exact age, sex, handedness, and pseudonymous ID, which must not be surfaced. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [ds005342 ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) — Idle & command · [idle](https://bci.report/protocols/idle/#foundation-v9) — OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — README reports voluntary signed consent and Universidad Antonio Nariño ethics approval. participants.tsv links pseudonymous IDs to exact age, sex, handedness, and academic major. The current evaluated subset has only four people. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, test-trial counts of idle false activations (of 60) and commands detected within 3 s (of 60), how many of the 4 people always abstained, and the cohort-mean window balanced accuracy and AUROC. No interval, no per-person or per-trial value. | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — SSVEP · 8 channels · [beta-8ch](https://bci.report/protocols/beta-8ch/#foundation-v9) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Paper reports Tsinghua University ethics approval 20190002 and written consent; parents consented for participants under 16. Ages span 9-64. Only cohort aggregates should be shown. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — SSVEP · 4 channels · [beta-4ch](https://bci.report/protocols/beta-4ch/#foundation-v9) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Paper reports Tsinghua University ethics approval 20190002 and written consent; parents consented for participants under 16. Ages span 9-64. Only cohort aggregates should be shown. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Arithmetic & rest · [arithmetic-rest](https://bci.report/protocols/arithmetic-rest/#foundation-v9) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate scores** — Paper reports Bioethics Commission approval (15 August 2018) and written consent from every participant. EDF dates are normalized, but subject-info.csv contains age, gender, occupation, and recording date; those fields must not be published. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. The EEGMAT adaptation rows add three-seed cohort means with participant-bootstrap intervals, per-seed cohort means, paired changes with how many of the 36 people were helped, harmed or tied, and trainable-parameter counts. | | [ds006593 ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) — P300 target ERP · [p300-target](https://bci.report/protocols/p300-target/#foundation-v9) — OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Dataset metadata/README report IRB approval #10-05-2022 and written informed consent. Participant metadata contains pseudonymous IDs and study descriptors without demographic fields. Eye-tracking and raw signals remain off-site. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [TMNRED / ds005383 ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) — Semantic target ERP · [semantic-target](https://bci.report/protocols/semantic-target/#foundation-v9) — Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0. | [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Paper reports Tianjin University ethics approval TJUE-2024-402, written consent, deidentification, and explicit participant permission for open sharing. Local participants.tsv appears column-shifted/malformed, so it must not feed public metadata. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [EESM19 scalp subset ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Sleep staging · [sleep-scalp](https://bci.report/protocols/sleep-scalp/#foundation-v9) — Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. | [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Paper reports Central Denmark Region ethics approval 1-10-72-413-17, Danish Medicines Agency approval 2017111085, registration, and written consent. Upstream includes exact age/sex plus sleep diaries, questionnaires, actigraphy, and timing; none should be exposed. Amended 2026-09-22: the 2025 data descriptor covering EESM19 and EESM23 (Mikkelsen et al., Scientific Data 12, 301, doi:10.1038/s41597-025-04579-8) states that publication was not mentioned in the informed consent form; before release, the GDPR office of Region Midt judged the data fully anonymized. Consent therefore covered the study, and the public release rests on that anonymization judgment. Reused unchanged from the 2026-09-20 review of the same recordings. Published here: for each new frozen encoder, the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | The new models’ weights, one entry per model family, with every checkpoint and its revision: | Model and checkpoints | Weights licence and terms | Owner’s review, 4 October 2026 | | --- | --- | --- | | [REVE Base · REVE Large](https://bci.report/methods/reve/) — REVE Base: `brain-bzh/reve-base @ dc2a075c` — REVE Large: `brain-bzh/reve-large @ 317531c7` | REVE Responsible Use License v1.0: aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; cite the REVE paper and name the model version; no adapted weights shared. — [Paper ↗](https://arxiv.org/abs/2510.21585) | — Accepted by the owner on 2026-10-04. Aggregate scientific results only; no re-identification, model inversion, membership inference or non-consensual profiling; the REVE paper is cited and the model version named; no adapted weights are shared. | | [LUNA Base · LUNA Large](https://bci.report/methods/luna/) — LUNA Base: `PulpBio/LUNA @ 999c1af0` — LUNA Large: `PulpBio/LUNA @ 999c1af0` | CC BY-ND 4.0 weights: scores from internal probing/adaptation only; no modified weights, LoRA or adapter deltas are shared; no implied endorsement by the LUNA authors; the LUNA name is not used to suggest an official release. — [Paper ↗](https://arxiv.org/abs/2510.22257) | — Owner-approved on 2026-10-04 for internal probing and adaptation with aggregate scores only. No modified weights, LoRA or adapter deltas are shared; no endorsement by the LUNA authors is implied, and the LUNA name is not used to suggest an official release. | | [BrainOmni Base](https://bci.report/methods/brainomni/) — BrainOmni Base: `OpenTSLab/BrainOmni @ 9a4d3c70` | MIT. — [Paper ↗](https://arxiv.org/abs/2505.18185) | — Permissive weights licence; aggregate scores only, no weights redistributed. | | [CodeBrain](https://bci.report/methods/codebrain/) — CodeBrain: `YjMajy/CodeBrain @ bef08d2f` | Apache-2.0. — [Paper ↗](https://arxiv.org/abs/2506.09110) | — Permissive weights licence; aggregate scores only, no weights redistributed. | | [EEGMamba](https://bci.report/methods/eegmamba/) — EEGMamba: `weighting666/EEGMamba @ 0b060d87` | MIT. — [Paper ↗](https://doi.org/10.1016/j.neunet.2025.107816) | — Permissive weights licence; aggregate scores only, no weights redistributed. | | [ST-EEGFormer Base · ST-EEGFormer Large](https://bci.report/methods/st-eegformer/) — ST-EEGFormer Base: `official GitHub release checkpoint-288.pth (code @ 542ee17)` — ST-EEGFormer Large: `official GitHub release large_weights_only_196.pth` | MIT (repository LICENSE; the README reserves the paper, diagrams and the ST-EEGFormer name). — [Paper ↗](https://openreview.net/forum?id=5Xwm8e6vbh) | — MIT repository licence; its README reserves the paper, the diagrams and the ST-EEGFormer name, none of which is reproduced. Aggregate scores only. | | [eeg-fm-masking MAE (r=9 cm, L=2) · eeg-fm-masking JEPA (r=9 cm, L=2) · eeg-fm-masking MAE (r=one channel, L=1) · eeg-fm-masking JEPA (r=one channel, L=1)](https://bci.report/methods/eeg-fm-masking/) — eeg-fm-masking MAE (r=9 cm, L=2): `PierreGtch/eeg-fm-masking_mae_r9cm_L2 @ c861ade2` — eeg-fm-masking JEPA (r=9 cm, L=2): revision not in the release — eeg-fm-masking MAE (r=one channel, L=1): revision not in the release — eeg-fm-masking JEPA (r=one channel, L=1): revision not in the release | CC-BY-4.0 weights: attribution required (cite arXiv:2609.33487). — [Paper ↗](https://arxiv.org/abs/2609.33487) | — Attribution licence; the paper is cited beside every row. Aggregate scores only. | | [ERP-FM Base](https://bci.report/methods/erp-fm/) — ERP-FM Base: `Google Drive file 1IXYnCOpDmY6raIn0M8_ggBOVey_A27l1 (no revision id; sha256 is the identity)` · SHA-256 `856564c4208e9cfe30cfeb5052038071c4951b7078de2bc9e461dead7eb511d4` | CC BY-NC-SA 4.0: non-commercial; aggregate scores only; attribution required; no weights redistributed. — [Paper ↗](https://arxiv.org/abs/2609.32796) | — Non-commercial share-alike weights. This site is personal noncommercial research and publishes aggregate scores only, with attribution; no weights are redistributed. | | [SingLEM](https://bci.report/methods/singlem/) — SingLEM: `ttlabtuat/SingLEM @ 6db5b5af` | MIT. — [Paper ↗](https://arxiv.org/abs/2509.17920) | — Checkpoint committed in the MIT repository; aggregate scores only. | | [ZUNA 1.1](https://bci.report/methods/zuna/) — ZUNA 1.1: `Zyphra/ZUNA1.1 @ 10b9d6c8` | Apache-2.0 weights; model card states research use only, not for diagnosis or clinical use. — [Paper ↗](https://arxiv.org/abs/2607.27308) | — Apache-2.0 weights. The model card says research use only, not for diagnosis or clinical use; that sentence travels with its rows. | ## Reviewed 7 October 2026 Route 2 of the decision-research roadmap: one representation and several questions. One new source, BOAS, and two records reused. OpenBMI keeps its rights record from the 3 October review, and EESM19 its record from the 20 September review with the 22 September amendment; route 2 uses every stored epoch of the full EESM19 release that passes the published quality rules, with the first scorer’s stages at their natural mix: a different protocol from the balanced scalp subset in the core matrix. Not published: per-window, per-person, per-night and per-fold values, including the per-person percentiles the release candidate carries for OpenBMI; measured compute totals (GPU-hours, wall time per fit and per block, training steps per outer fold), while the cost ledger publishes, labelled indicative, the two times measured interleaved on a shared GPU; and the aggregates and the three independent audits, which carry private storage paths and are pinned by hash in the review manifest instead. Release [`shared-representation-update-20261007`](https://bci.report/releases/#shared-representation-update-20261007). | Dataset | Source terms | What is published | | --- | --- | --- | | [OpenBMI motor imagery (Lee et al. 2019) ↗](https://doi.org/10.5524/100542) — Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — The paper's Ethical Approval section states that the study was approved by the Korea University Institutional Review Board (1040548-KUIRB-16-159-A-2) and that written informed consent was obtained from all participants before the experiments. Published here: participant-equal means with whole-participant bootstrap intervals for each arm and calibration budget, paired contrasts, and how many of the 51 people declined. The release's per-person summaries (worst observed change, medians, 10th percentiles) are dropped, and no participant, session file or trial identifier is published. Reused unchanged from the 2026-10-03 review of the same recordings. Published here for route 2: participant-equal means with person-level bootstrap intervals for each arm and question, cohort-level AUROC and log loss, paired contrasts with their intervals and per-seed points, cohort-level window counts and question prevalences, and the canary's cohort-level balanced accuracy on the first fold's 40 training people. The per-person 10th, 50th and 90th percentiles route 2 computed are dropped, as the 2026-10-03 release dropped its own. | | [BOAS, the Bitbrain Open Access Sleep dataset ↗](https://openneuro.org/datasets/ds005555/versions/1.1.3) — Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Released on OpenNeuro under CC0 by its authors (ds005555, version 1.1.3). The dataset description states approval by the Comité de Ética de la Investigación de la Comunidad Autónoma de Aragón (C.I. PI24/046). The README states that participants, adult members of the general population, provided written informed consent. Neither says whether the consent covered public sharing. Both statements come from the publisher's dataset description and README; no peer-reviewed paper describes BOAS. The ethics reference was added in version 1.1.1 (May 2025) with no approval date. Published here: cohort-level participant-equal means with person-level bootstrap intervals, paired contrasts with their intervals and per-seed points, log R with its interval, cohort-level AUROC, log loss and Cohen's kappa, cohort-level stage counts and question prevalences, the share of test windows where two heads disagree, and the canary's cohort-level balanced accuracy on the first fold's 80 training people; every cell pools at least 20 people (the smallest, 71). No per-person, per-night or per-fold value, no per-person percentiles, no identifier, date or demographic. Only the polysomnography EEG and the human-consensus labels are used. | | [EESM19, full release (first scorer) ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. | [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Paper reports Central Denmark Region ethics approval 1-10-72-413-17, Danish Medicines Agency approval 2017111085, registration, and written consent. Upstream includes exact age/sex plus sleep diaries, questionnaires, actigraphy, and timing; none should be exposed. Amended 2026-09-22: the 2025 data descriptor covering EESM19 and EESM23 (Mikkelsen et al., Scientific Data 12, 301, doi:10.1038/s41597-025-04579-8) states that publication was not mentioned in the informed consent form; before release, the GDPR office of Region Midt judged the data fully anonymized. Consent therefore covered the study, and the public release rests on that anonymization judgment. Reused unchanged from the 2026-09-20 review of the same release, with its 2026-09-22 amendment. Published here for route 2, in a panel labelled crude: every stored epoch of the first scorer that passes the published core quality rules, at the natural stage mix (a different protocol from the balanced scalp subset), and for the three questions participant-equal means of each arm, paired contrasts with person-level bootstrap intervals, cohort-level epoch counts and the canary's cohort-level balanced accuracy on the first fold's 16 training people. One seed; no per-person, per-night or per-fold value. | **BOAS, published with three stated gaps (owner decision, 7 October 2026).** Its authors released it on OpenNeuro under CC0, name an ethics committee with a reference number and state written informed consent. Three gaps remain, and they are stated wherever a BOAS figure appears: The consent statement does not say whether participants agreed to public sharing or secondary use. The ethics and consent statements come from the publisher's dataset description and README. No peer-reviewed paper describes BOAS. The ethics reference was added to the release in version 1.1.1 (May 2025), and the release does not say when it was granted relative to the recordings. The owner approved publishing cohort aggregates only, with every cell pooling at least 20 people, no demographics and no per-person distributions. Participants are pseudonymised in the public release. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. Credit: Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. ## Later sessions: three sources Four results on one question — does a decoder trained on an earlier session still work later? — each from its own approved release, and each the same person in a later session with no labels from it. The three sources, all published as aggregate scores, are WBCIC-SHU motor imagery, the longitudinal RSVP dataset (its Group A only) and Forenzo and He’s continuous-tracking dataset. Each licence was read from the repository’s public API record and each ethics and consent statement from the paper’s full text, on 8 October 2026. The continuous-tracking results count cohort records, which are not proven to be unique people, and keep the two cohorts apart. Not published: per-person and per-record values, including every minimum and percentile and every median other than the four cursor-tracking ridge medians printed beside their means; WBCIC-SHU confusion matrices; the cursor-tracking results split by recorded decoder; measured compute; and the handoffs, release decisions and independent audits, which carry private storage paths and are pinned by hash in the review manifest instead. Release [`later-sessions-update-20261008`](https://bci.report/releases/#later-sessions-update-20261008). | Dataset | Source terms | What is published | | --- | --- | --- | | [WBCIC-SHU motor imagery (Yang et al. 2025) ↗](https://doi.org/10.25452/figshare.plus.22671172.v5) — Banghua Yang, Fenqi Rong, Yunlong Xie, Du Li, Jiayang Zhang, Fu Li, Guangming Shi and Xiaorong Gao · A multi-day and high-quality EEG dataset for motor imagery brain-computer interface, Scientific Data 12, 488 (2025), doi:10.1038/s41597-025-04826-y. Data: Banghua Yang and Fenqi Rong, WBCIC-SHU Motor Imagery Dataset, Figshare+, doi:10.25452/figshare.plus.22671172.v5, CC BY 4.0. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The data paper states that written informed consent was obtained after the participants were informed about the procedure, and that the study was approved by the Tsinghua University Medical Ethics Committee (approval number 20190002) and adhered to the Declaration of Helsinki. Published here: participant-equal means with paired whole-participant bootstrap intervals for each cohort, the paired differences, secondary accuracy and macro-F1 means, and trial counts. No per-person value, participant or session-file identifier, confusion matrix or demographic is published. | | [Longitudinal ERP dataset, RSVP face task (Yang et al. 2025) ↗](https://doi.org/10.6084/m9.figshare.27201003.v1) — Chen Yang, Yufeng Zhang, Hongxin Zhang, Yixuan Li, Yijun Wang and Xiaorong Gao · A longitudinal EEG dataset of event-related potential, figshare, doi:10.6084/m9.figshare.27201003.v1, CC0. Paper: Yufeng Zhang, Hongxin Zhang, Yixuan Li, Yijun Wang, Xiaorong Gao and Chen Yang, A longitudinal EEG dataset of event-related potential, Scientific Data 12, 1069 (2025), doi:10.1038/s41597-025-05378-x. Derived analysis by BCI Report; not endorsed by the authors. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — The data paper states that the experiment was approved by the Tsinghua Institutional Review Board (No. 20230058) and that each participant read and signed an informed consent form before the experiment. Published here: participant-equal visit means for the 15 Group A participants, the AUROC interval at each visit, the paired Day 200 minus Day 7 contrast with its interval, how many participants declined by 0.05 AUROC or more, and event counts per visit. No per-person value, participant identifier, prediction or demographic is published. | | [Forenzo & He continuous-tracking EEG-BCI dataset (KiltHub) ↗](https://doi.org/10.1184/R1/25360300.v1) — Dylan Forenzo and Bin He · EEG-BCI Dataset for “Continuous Tracking using Deep Learning-based Decoding for Non-invasive Brain-Computer Interface”, KiltHub, Carnegie Mellon University, doi:10.1184/R1/25360300.v1, CC BY 4.0. Citation the record requests: Dylan Forenzo, Hao Zhu, Jenn Shanahan, Jaehyun Lim and Bin He, Continuous tracking using deep learning-based decoding for noninvasive brain–computer interface, PNAS Nexus 3(4), pgae145 (2024), doi:10.1093/pnasnexus/pgae145. Derived analysis by BCI Report; not endorsed by the authors and not a reproduction of the paper's online results. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The associated paper states that the study was approved by the Institutional Review Board at Carnegie Mellon University and that each subject gave written consent to the protocol before participating; it gives no approval number. Published here: for each cohort and response arm, record-equal means of the primary error for the spectral ridge and the source-mean comparator with bootstrap intervals, the ridge's median, the paired difference with its interval and how many admitted records the ridge did worse on, secondary means with undefined values kept null, and coverage counts. No other per-record value, no record, file or participant identifier, prediction or demographic is published. | **Ethics and consent, as each paper states them.** WBCIC-SHU motor imagery (Yang et al. 2025) — ethics approval: Approved by the Tsinghua University Medical Ethics Committee (approval number 20190002); carried out in line with the Declaration of Helsinki. Informed consent: Written informed consent was obtained from the participants after they were informed about the procedure, purpose, requirements and motor-imagery techniques. Read from [its paper’s full text ↗](https://europepmc.org/article/PMC/PMC11930978). Longitudinal ERP dataset, RSVP face task (Yang et al. 2025) — ethics approval: Approved by the Tsinghua Institutional Review Board (No. 20230058). Informed consent: Each participant read and signed an informed consent form before the experiment. Read from [its paper’s full text ↗](https://europepmc.org/article/PMC/PMC12185715). Forenzo & He continuous-tracking EEG-BCI dataset (KiltHub) — ethics approval: Approved by the Institutional Review Board at Carnegie Mellon University; no approval number is given. Informed consent: Each subject provided written consent to the protocol before participating. Read from [its paper’s full text ↗](https://europepmc.org/article/PMC/PMC11060102). ## Questions in language: sources, text encoders and wordings Route 3 of the decision-research roadmap, where one small head over frozen EEG features is asked each question by a question number, a label template or a description. No new dataset: BETA, BOAS, EESM19, OpenBMI motor imagery and the wearable SSVEP dataset keep their rights records, reused unchanged. New here are three frozen text encoders, which turn the wordings into vectors, and the wordings themselves. Not published: per-person, per-night, per-recording and per-fold values, per-window scores, features, head weights, text-encoder files and fold assignments; the per-frequency and interior-only SSVEP breakdowns, which no independent audit covers; any value of a pre-run check, which is published as pass or fail and counts only; demographics, recording dates and clock times; measured compute; and the handoff, the aggregates, the protocol and the four independent audits, which carry private storage paths and are pinned by hash in the review manifest instead. Release [`questions-in-language-update-20261008`](https://bci.report/releases/#questions-in-language-update-20261008). | Dataset | Source terms | What is published | | --- | --- | --- | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Cohort aggregates over the 70 people, with the published 2-s windows and participant folds: paired contrasts with their intervals and both flags, per-arm levels with chance-corrected values, generalisation costs, flip rates and the per-rotation TPL - CCA rows. No per-person, per-fold or per-frequency value. | | [BOAS, the Bitbrain Open Access Sleep dataset ↗](https://openneuro.org/datasets/ds005555/versions/1.1.3) — Eduardo López-Larraz, María Sierra-Torralba, Sergio Clemente, Galit Fierro, David Oriol, Javier Minguez, Luis Montesano and Jens G. Klinzing · The Bitbrain Open Access Sleep (BOAS) dataset, OpenNeuro ds005555, version 1.1.3 (2026), doi:10.18112/openneuro.ds005555.v1.1.3. Polysomnography EEG and the human-consensus stage labels only; the headband recordings and the publisher's automatic labels are not used. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Cohort aggregates only, each stating its n: 100 people, or 71 where the N3 stage needs at least five N3 windows per person (the per-stage AUROC for N3 and the "not N3" negation cell). No per-person, per-night, per-recording or per-fold value, no demographic, date or clock time; every pre-run check on BOAS as pass or fail and counts only. Read from route 2's frozen CBraMod and REVE-L features. | | [EESM19 ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. | [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Cohort aggregates over 20 people (S1, a replication of the primary sleep entries), each with its flags; declared likely inconclusive before the run. | | [OpenBMI motor imagery ↗](https://doi.org/10.5524/100542) — Min-Ho Lee, O-Yeon Kwon, Yong-Jeong Kim, Hong-Kyung Kim, Young-Eun Lee, John Williamson, Siamac Fazli and Seong-Whan Lee · EEG dataset and OpenBMI toolbox for three BCI paradigms: an investigation into BCI illiteracy, GigaScience (2019), giz002, doi:10.1093/gigascience/giz002. Data: Supporting data, GigaScience Database, doi:10.5524/100542. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Cohort aggregates over the published 51-person cohort (S2), each with its flags. Not gated by a triviality canary: none is declared for motor imagery. | | [Wearable SSVEP BCI dataset (dry and wet electrodes) ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) — Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 — Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces. Sensors, 21(4), 1256. https://doi.org/10.3390/s21041256 | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Cohort aggregates over 102 people for each sensor type (S12), computed from the official Tsinghua author mirror acquired on 2026-09-20; no physical unit is claimed and no participant row is published. | **Three text encoders, frozen at pinned commits and never redistributed.** [BAAI/bge-small-en-v1.5 ↗](https://huggingface.co/BAAI/bge-small-en-v1.5) (MIT, commit `5c38ec7c405ec4b44b94cc5a9bb96e735b38267a`); [sentence-transformers/all-MiniLM-L6-v2 ↗](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) (Apache-2.0, commit `1110a243fdf4706b3f48f1d95db1a4f5529b4d41`); [intfloat/multilingual-e5-small ↗](https://huggingface.co/intfloat/multilingual-e5-small) (MIT, commit `614241f622f53c4eeff9890bdc4f31cfecc418b3`). Each is used as released. **REVE Large, for secondary results.** Frozen REVE Large features (`brain-bzh/reve-large @ 317531c7`) feed secondary results on BETA and BOAS, under the REVE Responsible Use License v1.0, accepted by the owner with the other foundation models above: aggregate results only, the [REVE paper ↗](https://arxiv.org/abs/2510.21585) cited and the model version named. **The wordings are this site’s own text.** [questions-in-language-wordings.json](https://bci.report/data/questions-in-language-wordings.json) holds every text the questions were asked with, in English and Chinese, written by two agents from a fixed brief before any score, with the held-out frequencies and the derangements of the shuffled control. It is published under [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/). **BOAS, extended to route 3 (owner decision D11).** The owner extended BOAS’s approval of 7 October 2026 to route 3’s aggregate BOAS results, under the same conditions: the three stated gaps and the credit above, stated wherever a BOAS figure appears; participants pseudonymised in the public release; cohort aggregates only, every cell pooling at least 20 people; no demographics and no per-person distributions; every pre-run check on BOAS published as pass or fail and counts only. No result here evaluates Bitbrain's headband or its automatic sleep scoring; neither is used. ## Beyond scalp EEG: five sources Locally executed analyses beyond scalp EEG, each its own species, signal and metric, read on its own and outside the core matrix. Three sources have results — human fNIRS from Bak et al., one macaque’s motor-cortex spiking from MC_RTT, and a numerical coverage check of the ds007420 fNIRS recordings — and two are prepared and published as status only, with counts and no score: FALCON H2 human intracortical handwriting and Rabut et al. human functional ultrasound. Each licence was read from the repository’s public record, and each ethics and consent statement from the source named beside it. Bak’s statement comes from an Internet Archive capture of the publisher’s article page, because the publisher refused automated requests; nothing was bypassed. Rabut et al.’s data are licensed for non-commercial use only. Not published: per-person, per-recording, per-channel, per-trial, per-fold and per-row values; confusion matrices; signals, features, predictions, coefficients, labels and timestamps; measured compute and storage; demographics; and the handoff, its reviews, the protocols and the independent audits, which carry private storage paths and server names and are pinned by hash in the review manifest instead. Release [`beyond-scalp-update-20261012`](https://bci.report/releases/#beyond-scalp-update-20261012). | Dataset | Source terms | What is published | | --- | --- | --- | | [Bak 2019 fNIRS dataset · unilateral finger and foot tapping ↗](https://doi.org/10.6084/m9.figshare.9783755.v2) — human · fNIRS — Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong · Open Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, figshare, doi:10.6084/m9.figshare.9783755.v2, CC BY 4.0. Paper: Sujin Bak, Jinwoo Park, Jaeyoung Shin and Jichai Jeong, Open-Access fNIRS Dataset for Classification of Unilateral Finger- and Foot-Tapping, Electronics 8(12), 1486 (2019), doi:10.3390/electronics8121486. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The data paper states that the participants' written consent, approved by the Korea University Institutional Review Board (KUIRB-2019-0254-01), was submitted before data recording, and that the experiments followed the Declaration of Helsinki. Published here: the participant-macro balanced accuracy of a fixed classifier and of a training-prior comparator over the 30 people, the classifier's descriptive participant-bootstrap interval, and class recalls over all held-out predictions. No per-person value, participant identifier, confusion matrix or demographic is published. | | [MC_RTT · macaque motor cortex during self-paced reaching (Neural Latents Benchmark) ↗](https://doi.org/10.48324/dandi.000129/0.241017.1444) — macaque (Macaca mulatta) · intracortical spiking — O'Doherty, Joseph (2024) MC_RTT: macaque motor cortex spiking activity during self-paced reaching (Version 0.241017.1444) [Data set]. DANDI archive. https://doi.org/10.48324/dandi.000129/0.241017.1444, CC BY 4.0. Recorded by Joseph E. O'Doherty, Mariana M. B. Cardoso, Joseph G. Makin and Philip N. Sabes in the Sabes Lab at UCSF (Nonhuman Primate Reaching with Multichannel Sensorimotor Cortex Electrophysiology, Zenodo, doi:10.5281/zenodo.3854034; procedure: Makin et al., Journal of Neural Engineering 15(2), 026010 (2018), doi:10.1088/1741-2552/aa9e95). Prepared for the Neural Latents Benchmark: Felix Pei, Joel Ye et al., Neural Latents Benchmark '21: Evaluating latent variable models of neural population activity, NeurIPS 2021 Track on Datasets and Benchmarks, arXiv:2109.04463. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Animal data: one rhesus macaque. The Neural Latents Benchmark paper states that all its datasets were collected under procedures approved by the Institutional Animal Care and Use Committees at the respective institutions, and that MC_RTT was collected in the Sabes Lab at UCSF. No human took part, so informed consent does not apply. Published here: five validation metrics of a fixed ridge decoder and of a training-mean comparator, with trial, neuron and row counts. With one animal, every figure is that animal's own. No coefficient, prediction or per-trial value is published. | | [ds007420 · multi-distance fNIRS during ball squeezing and motion-artifact tasks ↗](https://doi.org/10.18112/openneuro.ds007420.v1.0.5) — human · fNIRS — Shibo Zhou, De’Ja Rogers, David A. Boas, Meryem A. Yücel and Yuanyuan Gao · Open-access Multi-Distance fNIRS Dataset Generated During Ball-Squeezing and Motion-Artifact Induction Tasks, OpenNeuro, doi:10.18112/openneuro.ds007420.v1.0.5, CC0. The dataset asks to acknowledge: Gao, Y., Rogers, D., von Lühmann, A., Ortega-Martinez, A., Boas, D.A., & Yücel, M.A. (2023). Short-separation regression incorporated diffuse optical tomography image reconstruction modeling for high-density functional near-infrared spectroscopy. Neurophotonics, 10(2), 025007. https://doi.org/10.1117/1.NPh.10.2.025007. Derived analysis by BCI Report; not endorsed by the authors. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Counts only — no score** — The dataset description names the Institutional Review Board of Boston University (Protocol #4502), and the associated paper states that the study protocol was approved by that board and that all participants provided written informed consent. Published here: whole-dataset counts from a label-independent numerical check of the optical values. No decoding score, per-recording or per-channel value, participant identifier or demographic is published. | | [FALCON H2 · human intracortical handwriting ↗](https://doi.org/10.48324/dandi.000950/0.241029.1403) — human · intracortical spiking — Fan, Chaofei; Hahn, Nick; Kamdar, Foram; Avansino, Donald; Wilson, Guy; Hochberg, Leigh; Shenoy, Krishna V; Henderson, Jaime; Willett, Frank (2024) FALCON Benchmark H2: Human Handwriting iBCI (Version 0.241029.1403) [Data set]. DANDI archive. https://doi.org/10.48324/dandi.000950/0.241029.1403, CC BY 4.0. Described in: Plug-and-Play Stability for Intracortical Brain-Computer Interfaces: A One-Year Demonstration of Seamless Brain-to-Text Communication, NeurIPS 2023. Part of the FALCON benchmark. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Status only — no scores** — The dataset description states that the participant gave informed consent and was enrolled in the BrainGate2 pilot clinical trial (ClinicalTrials.gov NCT00912041) under an FDA Investigational Device Exemption, with permission from the Institutional Review Boards of Stanford University (protocol #20804). Published here: status only, with trial, file and symbol counts from the preparation audit. No score exists; no neural value, cue text, recording date or participant detail is published. | | [Rabut et al. human functional ultrasound through a cranial window ↗](https://doi.org/10.22002/f3y3k-em558) — human · functional ultrasound imaging — Claire Rabut, Sumner Norman, Whitney Griggs, Kay Jann, Jonathan Russin, Vasileios Christopoulos, Charles Liu, Richard A. Andersen and Mikhail G. Shapiro · Dataset for: A window to the brain: ultrasound imaging of human neural activity through an acoustically transparent cranial prosthetic, CaltechData, doi:10.22002/f3y3k-em558, CC BY-NC 4.0. Paper: Claire Rabut, Sumner L. Norman, Whitney S. Griggs, Jonathan J. Russin, Kay Jann, Vasileios Christopoulos, Charles Liu, Richard A. Andersen and Mikhail G. Shapiro, Functional ultrasound imaging of human brain activity through an acoustically transparent cranial window, Science Translational Medicine 16, eadj3143 (2024), doi:10.1126/scitranslmed.adj3143. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY-NC 4.0 ↗](https://creativecommons.org/licenses/by-nc/4.0/) | **Status only — no scores** — The paper states that all procedures of the human study were approved by the Institutional Review Boards of the University of Southern California, Caltech (IR19-0902) and Rancho Los Amigos National Rehabilitation Hospital, and that the participant's consent was obtained. Published here: status only, with run and frame counts from the label audit. No image, score or participant detail is published. The dataset licence is noncommercial. | **Ethics and consent, as each source states them.** Bak 2019 fNIRS dataset · unilateral finger and foot tapping — ethics approval: Approved by the Korea University Institutional Review Board (KUIRB-2019-0254-01); the experiments followed the criteria of the Declaration of Helsinki. Informed consent: The participants' written consent, approved by that board, was submitted before data recording. Read from [doi.org ↗](https://doi.org/10.3390/electronics8121486). ds007420 · multi-distance fNIRS during ball squeezing and motion-artifact tasks — ethics approval: Approved by the Institutional Review Board of Boston University (Protocol #4502, as the dataset description records it). Informed consent: All participants provided written informed consent to take part in the study. Read from [europepmc.org ↗](https://europepmc.org/article/PMC/PMC10203730). FALCON H2 · human intracortical handwriting — ethics approval: Permission from the Institutional Review Boards of Stanford University (protocol #20804); the BrainGate2 pilot trial ran under FDA Investigational Device Exemption #G090003 (ClinicalTrials.gov NCT00912041). Informed consent: The participant gave informed consent and was enrolled in the BrainGate2 trial. Read from [api.dandiarchive.org ↗](https://api.dandiarchive.org/api/dandisets/000950/versions/0.241029.1403/). Rabut et al. human functional ultrasound through a cranial window — ethics approval: Approved by the Institutional Review Boards of the University of Southern California, Caltech (IR19-0902) and Rancho Los Amigos National Rehabilitation Hospital. Informed consent: The participant's consent was obtained. Read from [www.vis.caltech.edu ↗](https://www.vis.caltech.edu/documents/28341/scitranslmed.adj3143.pdf). **Animal data: an animal-care statement in place of consent.** MC_RTT · macaque motor cortex during self-paced reaching (Neural Latents Benchmark) records one rhesus macaque. Animal-care approval, as the Neural Latents Benchmark paper states it: Collected under procedures and experiments approved by the Institutional Animal Care and Use Committees at the respective institutions; MC_RTT was collected in the Sabes Lab at UCSF. Informed consent: Not applicable: animal data, no human participant. The DANDI record’s own ethics field is empty. Read from [the paper ↗](https://arxiv.org/abs/2109.04463). Its credit, above, names the people who recorded it and the Neural Latents Benchmark that prepared it. ## Pilot studies: four sources and one model Small pilots on four sources, each from its own approved aggregate export and read on its own, outside the core matrix. The Orthosis error-detection dataset is one cohort published in three configurations and counted once; BINGO is a cross-person, cross-language inner-speech pilot near its prior; REFED is a stored-clip computational diagnostic; the handwriting dataset’s mask-aware EEG analysis, five people, is published with its rights gap stated, by the owner’s decision. Each licence was read from the source’s own record, and each ethics and consent statement from the source named beside it. Not published: per-person, per-field, per-folder, per-clip and per-record values; the Orthosis response-clean subset’s scores, of which only counts are published; individual permutation and label-shuffle replicates, of which the mean and observed range are published; confusion matrices; identifiers, labels, features, predictions and fitted parameters; measured compute; demographics; and the handoffs, approval records, QA records and independent audits, which carry private storage paths and are pinned by hash in the review manifest instead. Status only: Orthosis EMG, TinyMyo and EEG–EMG fusion; handwriting fNIRS and fusion; REFED physiological and fusion evaluation; and a meditation dataset, in which RR means heartbeat RR intervals, not respiratory rate. Release [`pilot-studies-update-20261012`](https://bci.report/releases/#pilot-studies-update-20261012). | Dataset | Source terms | What is published | | --- | --- | --- | | [Orthosis error-detection EEG and EMG dataset (Kueper et al. 2024) ↗](https://doi.org/10.5281/zenodo.8345429) — Niklas Kueper, Kartik Chari, Judith Bütefür, Julia Habenicht, Tobias Rossol, Su Kyoung Kim, Marc Tabie, Frank Kirchner and Elsa Andrea Kirchner · EEG and EMG dataset for the detection of errors introduced by an active orthosis device (IJCAI’23 CC6 Competition), Zenodo, doi:10.5281/zenodo.8345429, CC BY 4.0. Paper: Niklas Kueper et al., EEG and EMG dataset for the detection of errors introduced by an active orthosis device, Frontiers in Human Neuroscience 18 (2024), doi:10.3389/fnhum.2024.1304311. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The data paper states that the study was approved by the Ethics Committee of the Department of Computer Science and Applied Cognitive Science of the Faculty of Engineering at the University of Duisburg-Essen, with no approval number given, and that the participants gave written informed consent. Published here: means over the 8 people with person-bootstrap intervals for three configurations of one cohort, the paired EEGPT-minus-classical difference with its interval, permutation-control means with their observed ranges, and cohort counts. No per-person value, participant or run identifier, response-clean subset score or demographic is published. | | [BINGO cross-linguistic inner-speech EEG dataset (Greek and English) ↗](https://bingo-project.gr/datasets/) — BINGO (Brain Imagined-Speech Communication) project, CERTH · cross-linguistic inner-speech EEG dataset, https://bingo-project.gr/datasets/; no licence is stated by the source. Documentation: Fotis P. Kalaganis, Kostas Georgiadis, Maria Kyrou, Spiros Nikolopoulos and Ioannis Kompatsiaris, D3.1 – The BINGO Benchmarking Framework, version 2.0, public deliverable (2025), https://bingo-project.gr/download/397. Derived analysis by BCI Report; not endorsed by the authors. | No licence stated | **Aggregate scores** — The source states no licence, and the cross-language dataset’s own documentation has no ethics-approval or consent statement. The project’s public deliverable D3.1, version 3.0, reports that participants signed a consent form; it says so for the project’s NATO-alphabet recordings in the same CERTH laboratory. No ethics-approval reference is published. Aggregates are published on that basis by the owner’s decision. Published here: means over the 17 source-defined subject fields with descriptive intervals, the cross-language contrast, controls and counts. No per-field value, field identifier, confusion matrix, prediction or source file is published, and no right to redistribute the data is implied. | | [EEG-fNIRS-based Handwriting Trajectory Dataset (Hugging Face) ↗](https://huggingface.co/datasets/lasfk/EEG-fNIRS-based-Handwriting-Trajectory-Dataset) — lasfk on Hugging Face · EEG-fNIRS-based Handwriting Trajectory Dataset, revision 4036680b5fb0015472d55589da2f34c5a87a3c10, CC BY-NC-SA 4.0 as declared on the dataset card, gated access. Derived analysis by BCI Report; not endorsed by the uploader. | [CC BY-NC-SA 4.0 ↗](https://creativecommons.org/licenses/by-nc-sa/4.0/) | **Aggregate scores** — No ethics-approval or informed-consent statement was found; the aggregates are published with that gap stated, by the owner’s decision. Published here: means over the 5 people with descriptive person-bootstrap intervals (none for the mask-only control, whose bounds would show individual people’s scores), three paired contrasts, label-shuffle means with observed ranges, the channel-availability distribution over trials and cohort counts. No per-person value, identifier, label, prediction or source file is published, and no right to redistribute the data is implied. | | [REFED: real-time dynamic labelled EEG-fNIRS emotion dataset (Ning et al. 2025) ↗](https://huggingface.co/datasets/REFED2025/REFED-dataset) — Xiaojun Ning, Jing Wang, Zhiyang Feng, Tianzuo Xin, Shuo Zhang, Shaoqi Zhang, Zheng Lian, Yi Ding, Youfang Lin and Ziyu Jia · REFED: A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset, NeurIPS 2025 Datasets and Benchmarks Track. Data: REFED2025/REFED-dataset on Hugging Face, revision 0298355c2823946ac5800b2ec12a082bf2dcd01b, CC BY-NC-SA 4.0, gated. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY-NC-SA 4.0 ↗](https://creativecommons.org/licenses/by-nc-sa/4.0/) | **Aggregate scores** — The paper states that the local Institutional Review Board approved the study (its datasheet names IRB-CASIA) and that all participants gave informed consent to the recording of their physiological signals and ratings, for non-commercial research use. Published here: equal-axis mean absolute errors over the 30 stored clips for four fixed arms in two conditions, counts and the method. No per-folder, per-clip or per-video value, target, prediction or demographic is published. | **Ethics and consent, as each source states them.** Orthosis error-detection EEG and EMG dataset (Kueper et al. 2024) — ethics approval: Approved by the Ethics Committee of the Department of Computer Science and Applied Cognitive Science of the Faculty of Engineering at the University of Duisburg-Essen; no approval number is given. Informed consent: The participants gave written informed consent and were told they could stop at any time. Read from [europepmc.org ↗](https://europepmc.org/article/PMC/PMC10839100). REFED: real-time dynamic labelled EEG-fNIRS emotion dataset (Ning et al. 2025) — ethics approval: Approved by the local Institutional Review Board; the paper’s datasheet names the approval IRB-CASIA. Informed consent: All participants gave informed consent to the recording of their physiological signals and rating scales, and agreed that the data would be used only for non-commercial scientific research. Read from [papers.neurips.cc ↗](https://papers.neurips.cc/paper_files/paper/2025/file/2bf0ed7c35d9d84128e7f7c72ab76402-Paper-Datasets_and_Benchmarks_Track.pdf). **BINGO, published with its rights as they stand (owner decision, 12 October 2026).** The project’s datasets page, which the analysed file comes from, states no licence, terms, ethics or consent; it gives an access link only. The cross-language dataset’s own documentation, [D3.1 – The BINGO Benchmarking Framework, version 2.0 ↗](https://bingo-project.gr/download/397) (Public (PU)), describes the protocol and the preprocessing and has no ethics or consent statement. Consent is as the project’s public deliverable [D3.1 – The BINGO Benchmarking Framework, version 3.0 ↗](https://bingo-project.gr/download/402) reports it: “After signing a consent form, each participant attended three experimental sessions”. That sentence is about the project’s NATO-alphabet dataset, recorded in the same CERTH laboratory, not about the cross-language dataset. No ethics-approval reference is published by the source. The owner decided to publish the aggregates on that basis, with these rights stated wherever the result is. Nothing upstream is redistributed, and no right to reuse the data is implied. **The handwriting dataset, published with its rights gap stated (owner decision, 12 October 2026).** No ethics-approval or informed-consent statement was found. The dataset card has none; a participating team’s challenge paper, which identifies the competition as the ACM MM 2026 Multimodal Brain-Computer Interface Grand Challenge, has none; and the organisers’ paper is listed there as forthcoming. The aggregates are published with this gap stated, by the owner’s decision, to be rechecked when the organisers’ paper appears. The challenge paper read: [FRED: Frequency-Decorrelated Temporal Ensembles for EEG–fNIRS Imagined-Handwriting Decoding ↗](https://arxiv.org/abs/2608.03176v2). The dataset card declares CC BY-NC-SA 4.0 and gated access. Like BOAS and BINGO, this is an owner decision on stated gaps; the BNCI2015-001 hold is unchanged, because its review turns on its CC BY-NC-ND catalogue licence as well. Nothing upstream is redistributed, and no right to reuse the data is implied. **EEGPT in the pilot studies.** Frozen EEGPT with a fixed supervised probe, against matched classical features on the Orthosis error-detection dataset, is published from its approved export, after an independent check of that exact file. The earlier hold on EEGPT aggregate scores was this project’s own conservative choice, not a publisher’s rule. This comparison was the first EEGPT result on the site’s pages; its core-matrix cells and pretraining-attribution rows followed on 13 October 2026, in the section [EEGPT on the core matrix](https://bci.report/data-use/#eegpt-core) below. Some of them had been in audit records in the public repository from its first release, `research-preview-20260920`: balanced accuracy for mi-rest, EEGMAT and both BETA protocols (`research/benchmark_expansion_v1/`), and EEGPT’s pretraining-attribution controls on mi-rest and EEGMAT (`research/deployment_validation_20260920/`). Corrected 12 October 2026: this paragraph first said the comparison was the only EEGPT figure published. The code licence, Apache-2.0, covers the repository’s source code only; the checkpoint’s licence is unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights. No original or converted weights are distributed: the checkpoint is named and its SHA-256 given (`eegpt_mcae_58chs_4s_large4E.ckpt`, `9d63ebc80bba1ad5e3277df4eecfa05d2745e6243a12ff743c4c0a93930dfac0`) to identify it, and that grants nothing. Repository [BINE022/EEGPT ↗](https://github.com/BINE022/EEGPT) at `a0e0a8fad729e2ecf4eedb3a81548a6e6d48a705`; paper [EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals ↗](https://openreview.net/forum?id=XNiFuoqTuD). ## EEGPT on the core matrix By the owner’s decisions of 12 and 13 October 2026: Publish EEGPT’s eight core-matrix cells as computed, without a re-run, as a frozen row beside the core matrix, and its pretraining-attribution rows beside the published LaBraM and CBraMod rows. No new dataset: the cells are the recorded aggregates of the same runs as the published LaBraM rows and, except on idle, the CBraMod rows, with only the encoder swapped, checked read-only against the saved predictions. The seven datasets behind the protocols keep their rights records from the 2026-09-20 review, reused unchanged. Release [`eegpt-core-update-20261013`](https://bci.report/releases/#eegpt-core-update-20261013). | Dataset · protocol | Source terms | What is published | | --- | --- | --- | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Motor imagery & rest · [mi-rest](https://bci.report/protocols/mi-rest/#eegpt) — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [ds005342 ↗](https://doi.org/10.18112/openneuro.ds005342.v1.0.3) — Idle & command · [idle](https://bci.report/protocols/idle/#eegpt) — OpenNeuro ds005342 contributors · version 1.0.3, doi:10.18112/openneuro.ds005342.v1.0.3; associated study doi:10.3389/fninf.2022.961089. Original author credits are retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: test-trial counts of idle false activations and commands detected within 3 s, how many people always abstained, and the cohort-mean window balanced accuracy and AUROC; no interval; no per-person, per-fold or per-trial value. | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — SSVEP · 8 channels · [beta-8ch](https://bci.report/protocols/beta-8ch/#eegpt) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — SSVEP · 4 channels · [beta-4ch](https://bci.report/protocols/beta-4ch/#eegpt) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Arithmetic & rest · [arithmetic-rest](https://bci.report/protocols/arithmetic-rest/#eegpt) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [ds006593 ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) — P300 target ERP · [p300-target](https://bci.report/protocols/p300-target/#eegpt) — OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [TMNRED / ds005383 ↗](https://doi.org/10.18112/openneuro.ds005383.v1.0.0) — Semantic target ERP · [semantic-target](https://bci.report/protocols/semantic-target/#eegpt) — Yanru Bai, Qi Tang et al. · TMNRED, A Chinese Language EEG Dataset for Fuzzy Semantic Target Identification in Natural Reading Environments (2025), doi:10.1038/s41597-025-05036-2. OpenNeuro ds005383 v1.0.0. | [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | | [EESM19 scalp subset ↗](https://doi.org/10.18112/openneuro.ds005185.v1.0.2) — Sleep staging · [sleep-scalp](https://bci.report/protocols/sleep-scalp/#eegpt) — Kaare B. Mikkelsen et al. · Accurate whole-night sleep monitoring with dry-contact ear-EEG (2019), doi:10.1038/s41598-019-53115-3; OpenNeuro ds005185 v1.0.2. Processed mirror: Zachary1150/EESM19-Processed. | [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for EEGPT: the cohort balanced accuracy with its participant-bootstrap interval, macro F1 and the cohort size; no per-person, per-fold or per-trial value. | **The model’s terms.** Same reasoning as the 2026-10-12 pilot decision: independently computed aggregate results are a separate output from the weights; no publisher prohibition was found; the checkpoint’s licence is unresolved; no original or converted weights are distributed; Apache-2.0 covers the repository’s code only. The checkpoint is the one the pilot comparison above identifies by its SHA-256; its licence is unresolved: the checkpoint’s own licence and terms could not be read from its hosting record, and the repository’s code licence is not evidence that it covers the weights. No model’s authors endorse these results. **How EEGPT differs from EEG-DINO.** EEG-DINO stays catalogue-only by the owner’s decision of 2026-10-04: its upstream weights carry no licence, and it was neither downloaded nor run. EEGPT differs: it was evaluated in the same runs as the published rows and was held only because its checkpoint’s licence could not be read, a hold the 2026-10-12 reassessment found to be this project’s own choice. EEG-DINO is not re-reviewed by this decision. **The public repository’s records.** Four of the eight cells (mi-rest, EEGMAT and both BETA protocols) and EEGPT’s attribution controls have been in audit records in the public repository from its first release, `research-preview-20260920`; this release puts them on the site’s pages with the other four cells. The public attribution audit (research/deployment_validation_20260920/attribution-independent-audit.json) still labels the four EEGPT contrasts “hold-unresolved-weight-license”: that label records 2026-09-20. With this release they are published. The file keeps its bytes. **Pretraining exposure.** None of the seven core datasets is in the pretraining list EEGPT’s authors publish, checked on 2026-10-12; BETA’s relation to the Tsinghua benchmark that is in the list is stated beside the BETA cells. **Not published.** - Per-trial predictions and scores, per-person, per-fold and per-class metrics, features, fitted heads, idle thresholds and calibration grids, and every identifier. How many people had a negative attribution difference is not published either. - The per-seed balanced accuracies of the three constructor-random encoders and their spread: as for LaBraM and CBraMod, only their mean is published. - BETA’s descriptive duration-group summaries (S1–15 against S16–70): the published BETA rows show none. - The idle sensitivity counts of the review (how many windows or pairs lie near a threshold), its pooled AUROC, which is not the protocol’s definition, and the Monte Carlo spreads of the independent recomputation. - Timings, devices, memory and software environments. - EEGPT results outside the eight core protocols (BNCI2014-001 and the other parallel-batch datasets). - The checkpoint and its tensor conversion: no original or converted weights are distributed; their SHA-256 identifies them and grants nothing. - The run summaries, protocols and independent audits, the batch state and the review’s verification records themselves: each is pinned by SHA-256 in the release manifest and checked by the export, never published, because they carry private storage paths. ## Data quality: three sources checked end to end Three public EEG datasets, each read from its own files and read on its own, outside the core matrix and every leaderboard. Shin2022 is a new source, published as aggregate scores: a participant-held-out diagnostic with three arms, beside the coverage chain that counts its runs and windows. eldBETA is a new source, published in this release as counts for one participant; its file checks and cohort result for all its participants have their own release, below. ZuCo 2.0 is status only: a timing hold on one converted run, with no figure. Each licence was read from the source’s own record and each ethics and consent statement from the paper named beside it, and the export records each read with its capture’s SHA-256. Not published: participant rows of any kind: identifiers, run and file names, fold assignments, features, fitted models, decision values and predictions; raw signals from any of the three datasets; per-participant figures from the Shin2022 evidence: window counts per participant and their ranges, and which participants had flat windows; the Shin2022 feedback-duration distribution (quantiles and mean) and the per-channel counts of the signal census; eldBETA's one-participant accuracy and its fixed first-frequency comparator: held by the owner's decision under the one-person rule; ZuCo 2.0's converted-run code, row-level timing values and marker counts; no ZuCo 2.0 data are redistributed; MOABB's catalog trial and run counts for Shin2022: register content, in the open register file; compute times, storage sizes, hosts and storage locations; the research exports, audits, reviews, protocols and handoffs themselves: each is pinned by SHA-256 in the release manifest and checked by the export, never published, because several carry private storage paths. Release [`data-quality-update-20261020`](https://bci.report/releases/#data-quality-update-20261020). | Dataset | Source terms | What is published | | --- | --- | --- | | [Shin2022 · Data from: Closed-loop motor imagery EEG simulation for brain-computer interfaces (Shin, Suma and He) ↗](https://doi.org/10.6084/m9.figshare.20383716.v1) — Figshare record 20383716, version 1, the only version the record lists — Hyonyoung Shin, Daniel Suma and Bin He · Data from: Closed-loop motor imagery EEG simulation for brain-computer interfaces, figshare, doi:10.6084/m9.figshare.20383716.v1, CC BY 4.0. Paper: Hyonyoung Shin, Daniel Suma and Bin He, Closed-loop motor imagery EEG simulation for brain-computer interfaces, Frontiers in Human Neuroscience 16:951591 (2022), doi:10.3389/fnhum.2022.951591. Funding: NIH AT009263. Modifications: canonical deduplication of the byte-identical runs, first-second window selection and signal-validity filtering. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Aggregate scores** — The study was reviewed and approved by the Institutional Review Board of Carnegie Mellon University; the paper gives no approval number. The participants gave written informed consent to take part in the study. The data record states that all identifying information has been removed. Published here: each arm’s participant-mean balanced accuracy with its participant-bootstrap interval and pooled secondary value, and aggregate run and window counts; no per-participant value, identifier, prediction or signal. | | [eldBETA: A Large Eldercare-oriented Benchmark Database of SSVEP-BCI for the Aging Population (Liu, Wang, Gao and Chen) ↗](https://doi.org/10.6084/m9.figshare.18032669.v3) — Figshare record 18032669, version 3, the latest of the 3 versions the record lists — Bingchuan Liu, Yijun Wang, Xiaorong Gao and Xiaogang Chen · eldBETA: A Large Eldercare-oriented Benchmark Database of SSVEP-BCI for the Aging Population, figshare, doi:10.6084/m9.figshare.18032669.v3, CC BY 4.0. Data descriptor: Scientific Data, doi:10.1038/s41597-022-01372-9. Funding: National Natural Science Foundation of China (No. 62171473). Modifications: decoding of the original recording format with the header calibration, and fixed channel and window selection; no score from these data is published. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | **Counts for one participant in this release; its cohort result has its own release below** — The protocol followed the Declaration of Helsinki and was approved by the institutional review board of Tsinghua University, approval number 20210032. The participants were familiarised with the protocol and gave full written consent before the experiment. Published here: counts for one participant’s recordings (blocks, targets, trials, trials dropped and non-finite values) and no score; no identifier, per-trial value or signal. | | [ZuCo 2.0 ↗](https://data.nemar.org/nm000279/v1.0.0/) — NEMAR nm000279, version 1.0.0, doi:10.82901/nemar.nm000279.v1.0.0 — Hollenstein, N., Troendle, M., Zhang, C., & Langer, N. (2020). ZuCo 2.0: A Dataset of Physiological Recordings During Natural Reading and Annotation. LREC 2020. | CC-BY-4.0 — As the dataset_description.json of that version states it. Unverified: the original terms on OSF; terms for the text excerpts the stimuli quote. | **Status only — no score** — In one converted run on NEMAR, some fixation times fall outside the recording; the run’s stimulus markers lie within it. No EEG–text score is computed until the converter's timing rules and a source-checksummed fixture are available. The hold covers that one converted run; no timing was corrected and no row removed. | **Ethics and consent, as each source states them.** Shin2022 — ethics approval: The study was reviewed and approved by the Institutional Review Board of Carnegie Mellon University; the paper gives no approval number. Informed consent: The participants gave written informed consent to take part in the study. Read from [www.frontiersin.org ↗](https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2022.951591/full). eldBETA — ethics approval: The protocol followed the Declaration of Helsinki and was approved by the institutional review board of Tsinghua University, approval number 20210032. Informed consent: The participants were familiarised with the protocol and gave full written consent before the experiment. Read from [www.nature.com ↗](https://www.nature.com/articles/s41597-022-01372-9). ZuCo 2.0 — No consent or ethics statement for this dataset was read in this release. **Terms the records state beside the licence.** Shin2022: “All identifying information have been removed.” (Figshare record description) “Please cite the above paper if you use any data included in this dataset.” (Figshare record description) eldBETA: “The eldBETA database is open-access for research and can be downloaded from the website” (data descriptor, abstract) The papers carry their own licences, which cover the articles; each data licence is its record’s. Shin2022’s attribution names the modifications made here, and so does eldBETA’s. ## Clinical groups: IESSEEG An infant case-control pilot on one public dataset, published as aggregate scores in its own section of the clinical-groups page, with age and recording-setting controls and frozen NeuroLM probes. IESSEEG is a new source, and it adds no dataset page and no count. The clinical bar above applies to it: Research results on a public data set. Not a diagnosis, not diagnostic accuracy, and not a medical device or medical advice. The licence and the attribution come from the release card and LICENSE at the pinned revision, and the requested citation is the card’s: “IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome”. Not published: participant rows of any kind: identifiers, ages, diagnoses, recordings, fold assignments, features, out-of-fold predictions and probabilities; per-fold metrics for the five folds and for the age <=2 band within each fold; per-group age means, medians and quartiles, and sex counts: aggregates, but a demographic profile of a diagnosis-specific group, which this site does not publish. The age <=2 column gives only the band's size and its case and control counts; intervals and significance tests: none was specified before scoring, and none is computed after it; EEG signals, model code, the NeuroLM checkpoint and its features: none is redistributed; runtimes, hosts, storage locations and the private seal and score files (named by SHA-256 only). Release [`clinical-iesseeg-update-20261020`](https://bci.report/releases/#clinical-iesseeg-update-20261020). | Dataset | Source terms | What is published | | --- | --- | --- | | [IESSEEG · awake case-control pilot ↗](https://huggingface.co/datasets/Capur/IESSEEG/tree/1c32ee3a425a324808a9d275b60ce85625af35c8) — Hugging Face Capur/IESSEEG at revision 1c32ee3a425a324808a9d275b60ce85625af35c8 — Dataset: IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome, Hugging Face Capur/IESSEEG at revision 1c32ee3a, CC BY 4.0 for the dataset and annotations. The release card declares that the 400 case clinician-selected segments were previously analysed by Rajaraman et al. (2024), Computational EEG attributes predict response to therapy for epileptic spasms. | [CC-BY-4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — For the dataset and annotations; [LICENSE ↗](https://huggingface.co/datasets/Capur/IESSEEG/blob/1c32ee3a425a324808a9d275b60ce85625af35c8/LICENSE) | **Aggregate scores** — The release card states institutional review board approval of the retrospective study, with waivers of informed consent and HIPAA research authorization, and asks that no one attempt to re-identify participants. The publication boundary is the one the Parkinson’s comparison uses: aggregate scores only, with no participant rows, predictions, ages or per-group age and sex composition; the age ≤2 column gives the band’s size and its case and control counts. | **Ethics and consent, as the release card states them.** “The retrospective study received institutional review board approval under protocol 17-000026, with waivers of informed consent and HIPAA research authorization. Do not attempt to re-identify participants.” Read from [the release card at revision `1c32ee3a` ↗](https://huggingface.co/datasets/Capur/IESSEEG/blob/1c32ee3a425a324808a9d275b60ce85625af35c8/README.md). Informed consent and HIPAA research authorization were waived for this retrospective study, so the published aggregates rest on the board’s approval and the card’s licence. **Prior reuse, as the publisher declares it.** The release card declares that 400 case segments were previously analysed by Rajaraman et al. (2024), “Computational EEG attributes predict response to therapy for epileptic spasms”. The participant-level crosswalk between that analysis and this release is unresolved; the cohort adds no dataset to any count. **NeuroLM in this section.** NeuroLM-B is a frozen feature extractor read by supervised logistic probes; its weights are unchanged, and no weights are distributed. The checkpoint is the one the NeuroLM pilot below identifies by its SHA-256 (`ffe098bc138b89f8817d3710a3604498d8ecd15135080e2ca27735d05c6d29ab`), under the terms listed there. Its training stage and its prior exposure to these participants are unknown. ## NeuroLM on EEGMAT: a separate pilot panel NeuroLM-B’s frozen features on EEGMAT’s arithmetic and rest, published as aggregate scores in a panel of its own on the pilot-studies page, outside the pilot count and every leaderboard. No new dataset: EEGMAT’s rights record from the 2026-09-20 review is reused unchanged, with what this release publishes on it in the table below. EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. NeuroLM’s code is MIT at the pinned commit. The NeuroLM-B checkpoint is CC BY 4.0 as the licence tag of its Hugging Face repository states, and its 30-byte model card adds no terms. No weights are distributed, and no endorsement by NeuroLM’s authors is implied. Pretraining exposure is unknown. EEGMAT is NeuroLM’s Workload task, one of the six instruction tasks in its pinned code, and the training stage of this checkpoint is unresolved: its metadata (epoch 19, iteration 83,660, initialised from GPT-2, learning rate 0.0006) fits a pretraining checkpoint, and instruction tuning is unverified. The backbone may have seen these people, so exposure is shown as unknown. Not published: per-person, per-fold and per-window values of any kind, fold assignments, participant identifiers, condition labels, features, fitted heads, predictions and the sealed prediction files: only equal-person means, paired differences with their intervals and cohort counts are published; demographics from the source dataset’s participant file; measured compute: worker times, memory and process counts from the handoff, the receipts and the audits; the checkpoint and the NeuroLM source: no weights are distributed; the SHA-256 identifies the checkpoint used and grants nothing; the generation-mask adapter’s checks: engineering evidence that took no part in these scores; the IESSEEG intake and the watchlist leads of the same handoff: IESSEEG has its own release, and the leads have no results; the handoff, protocol, producer, receipts, reviews and input manifest themselves: each is pinned by SHA-256 in this manifest and checked by the export, and none is published, because several carry private storage paths. Release [`neurolm-eegmat-update-20261020`](https://bci.report/releases/#neurolm-eegmat-update-20261020). | Dataset | Source terms | What is published | | --- | --- | --- | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — PhysioNet EEG During Mental Arithmetic Tasks 1.0.0 · 1.0.0 — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for NeuroLM’s frozen features: equal-person balanced accuracy and macro-F1 of three representations, three paired differences with their participant-bootstrap intervals, and cohort counts; no per-person, per-fold or per-window value. | The model’s terms, part by part: | Part | Terms | Read from | | --- | --- | --- | | Code — [935963004/NeuroLM ↗](https://github.com/935963004/NeuroLM/tree/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177) at `0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177` | MIT — NeuroLM’s source code at the pinned commit | [LICENSE ↗](https://github.com/935963004/NeuroLM/blob/0cda9876d8ce6ee07ed0c43eee5e9a6f5c24b177/LICENSE) — SHA-256 `21bddd47eec5ad9a56bc9bec45f91426680e373f957b840fde984252b6718185` | | Checkpoint — [Weibang/NeuroLM ↗](https://huggingface.co/Weibang/NeuroLM/tree/eddfff5c64a4139442f826d6c67c8369fd00f45a) at `eddfff5c64a4139442f826d6c67c8369fd00f45a`: `checkpoints/NeuroLM-B.pt`, SHA-256 `ffe098bc138b89f8817d3710a3604498d8ecd15135080e2ca27735d05c6d29ab` | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — As the licence tag in the Hugging Face repository metadata at the pinned revision; its model card is 30 bytes and adds no terms. No weights are distributed; the SHA-256 identifies the checkpoint used and grants nothing. | Checked on 2026-10-16 | **The model, as its authors publish it.** [NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals ↗](https://arxiv.org/abs/2409.00101). NeuroLM is used here as a frozen feature extractor, and the transfer is supervised: the frozen NeuroLM outputs feed a linear head trained on the other folds’ labels. No endorsement by its authors is implied. ## New SSVEP frequencies: cards 8 to 14 Cards 8 to 14 of the new-SSVEP-frequencies page: follow-ups on SSVEP frequencies a decoder never trained on, each read from its own audited aggregate and published as aggregate results, outside every leaderboard. No new dataset: BETA’s and Wearable-102’s released records are reused unchanged, with what these cards publish on each in the table below. BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. Wearable-102 keeps its released record: CC BY 4.0, cohort aggregates only. Filter-bank CCA is credited to Chen et al. 2015; the code that ran is the project’s own SciPy realisation, and no code or weights are distributed. Not published: per-person, per-fold, per-rotation, per-seed and per-window values of any kind, predictions, scores, features, fitted heads and the sealed run files: only cohort levels, paired contrasts with their intervals, gap statistics, decision-change shares, check counts and CPU component medians are published; per-frequency values of any kind, including the reference check’s candidate frequencies: the check is published as counts; the cross-spectral per-seed MLP accuracies, the canary accuracies and the random-pairing accuracies of cards 8 and 9; the fixed-decay retrospective gap-to-gain ratio table and its dry reconstruction; all-trial decision counts: the harmonic-gap decision changes on BETA held-out are published as shares of the 11,200 decisions; the Wearable-102 decision changes and the fixed-decay and rising-weight transitions are not published; the harmonic-gap Delta means and the margins against the strongest joint-CCA distractor; run times of whole evaluations and audits, memory and process identifiers; the CPU-time readings of the component measurements and every host, processor, system and runtime label; predictions from the comment thread on cards 8 to 14: each new card states its prespecified criteria only, and the thread’s prediction stays on card 4 of the released file; the commenter’s name; the handoffs, protocols, freezes, receipts, reviews and approvals themselves: each is pinned by SHA-256 in the release manifest and checked by the export; several carry private storage paths. Release [`ssvep-followups-update-20261028`](https://bci.report/releases/#ssvep-followups-update-20261028). | Dataset | Source terms | What is published | | --- | --- | --- | | [BETA ↗](https://figshare.com/articles/dataset/The_BETA_database/12264401) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. — Rights reviewed on 2026-09-20; reused unchanged from the released record. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — Attribution required. Participant consent described use for non-commercial scientific research, a narrower operational boundary than the copyright license; minors required parental consent. Cohort aggregates only; no participant records, EEG signals, or individual predictions. | **Aggregate results** — Cohort aggregates over the 70 people for cards 8 to 14: levels and paired contrasts with their intervals, gap statistics in squared-correlation units, decision-change shares and check counts; and CPU component medians measured on BETA windows. No per-person, per-fold, per-rotation, per-window or per-frequency value. | | [Wearable-102 · An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces ↗](https://figshare.com/articles/dataset/An_Open_Dataset_for_Wearable_SSVEP-Based_Brain-Computer_Interfaces/13560281) — Zhu, F., Jiang, L., Dong, G., Gao, X., & Wang, Y. (2021). An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces (Version 4) [Data set]. Figshare. https://doi.org/10.6084/m9.figshare.13560281.v4 — Official Tsinghua BCI Lab author mirror snapshot acquired 2026-09-20; the mirror itself has no separate DOI or version number. | CC BY 4.0 — As the [Figshare record ↗](https://doi.org/10.6084/m9.figshare.13560281.v4) states it. The paper describes consent for noncommercial scientific research. | **Aggregate results** — Cohort aggregates over the 102 people, by sensor, for cards 10, 12, 13 and 14: levels and paired contrasts with their intervals, gap statistics and check counts. No per-person, per-window, per-frequency or questionnaire value. | **Filter-bank CCA.** Credited to [Chen X, Wang Y, Gao S, Jung T-P, Gao X. Filter bank canonical correlation analysis for implementing a high-speed SSVEP-based brain–computer interface. Journal of Neural Engineering 12(4):046008, 2015. ↗](https://doi.org/10.1088/1741-2560/12/4/046008) Used here as filter-bank CCA (seven Chebyshev sub-bands, five harmonics), our SciPy realisation of the published recipe; no code is distributed. Route 3’s questions-in-language page runs another filter-bank recipe on the same task, with 5 Butterworth sub-bands. ## NeuroRVQ on three tasks: a separate pilot panel NeuroRVQ’s released backbone, kept frozen and read by one fixed ridge head, on EEGMAT’s arithmetic and rest, on motor imagery against rest and on P300 target against non-target, published as aggregate scores in a panel of its own on the pilot-studies page, outside the pilot count and every leaderboard; the three protocol pages point to it with no figure. No new dataset: each task’s released rights record is reused unchanged, in the table below. NeuroRVQ’s code and weights are CC BY-NC 4.0, the licence file of both repositories at the pinned revisions; only aggregate figures are published, no weights are distributed, and no endorsement by NeuroRVQ’s authors is implied. EEGMAT, ds003810 and ds006593 keep their released records. Pretraining exposure is unknown: none of the three datasets is named in the paper’s list of 13 pretraining corpora, 1 of them proprietary (checked on 2026-10-25). Not published: per-person, per-fold and per-window values of any kind, fold assignments, participant identifiers, labels, extracted values, fitted heads, predictions and the sealed files: only participant-mean scores, one paired difference per task with its interval, and cohort counts are published; how many people each arm scored above 50%, which the sources give as descriptive point-estimate counts; per-arm intervals and macro-F1 differences: none were computed, so each arm has a point estimate only; measured compute: workflow, audit and CPU times, memory, budgets and process counts from the handoffs and receipts. None of them measures inference speed; the checkpoint and the authors’ source: no weights are distributed; the checkpoint SHA-256 identifies the file used and grants nothing. No hash of a source or Hugging Face file is published beyond the checkpoint SHA-256 and the two revisions; third-party conversions of the checkpoint and the licence tags they carry: only the authors’ repositories are cited; the handoffs, protocols, configurations, input queues, receipts and reviews themselves: each is pinned by SHA-256 in this manifest and checked by the export, and none is published, because several carry private storage paths. Release [`neurorvq-update-20261028`](https://bci.report/releases/#neurorvq-update-20261028). | Dataset | Source terms | What is published | | --- | --- | --- | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — PhysioNet EEG During Mental Arithmetic Tasks 1.0.0 · 1.0.0 — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) — Personal noncommercial research; aggregate results only | **Aggregate scores** — Reused unchanged from the 2026-09-20 review. Paper reports Bioethics Commission approval (15 August 2018) and written consent from every participant. EDF dates are normalized, but subject-info.csv contains age, gender, occupation, and recording date; those fields must not be published. | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — OpenNeuro snapshot 2.0.2 · c674303319303b9ffc4ff7e2340b3a4807e8ffb5 — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) — Personal noncommercial research; aggregate results only | **Aggregate scores** — Reused unchanged from the 2026-09-20 review. Study paper reports ethics approval by CEySTE CCT-CONICET Santa Fe, Declaration of Helsinki compliance, and signed consent. Twelve adults participated; ten were retained. Public participants.tsv includes exact age, sex, handedness, and pseudonymous ID, which must not be surfaced. | | [ds006593 ↗](https://doi.org/10.18112/openneuro.ds006593.v1.0.0) — OpenNeuro snapshot 1.0.0 · b3fa345b310abfcf442dd4e2867f6652336c3fe4 — OpenNeuro ds006593 contributors · version 1.0.0, doi:10.18112/openneuro.ds006593.v1.0.0; original author credits retained at the linked source. | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) — Personal noncommercial research; aggregate results only | **Aggregate scores** — Reused unchanged from the 2026-09-20 review. Dataset metadata/README report IRB approval #10-05-2022 and written informed consent. Participant metadata contains pseudonymous IDs and study descriptors without demographic fields. Eye-tracking and raw signals remain off-site. | The model’s terms, part by part: | Part | Terms | Read from | | --- | --- | --- | | Code — [KonstantinosBarmpas/NeuroRVQ ↗](https://github.com/KonstantinosBarmpas/NeuroRVQ/tree/926e770d9d16b6aa308404280fa0cc0211a6f9fb) at `926e770d9d16b6aa308404280fa0cc0211a6f9fb` | [CC BY-NC 4.0 ↗](https://creativecommons.org/licenses/by-nc/4.0/) — NeuroRVQ’s source code at the pinned revision | [LICENSE ↗](https://github.com/KonstantinosBarmpas/NeuroRVQ/blob/926e770d9d16b6aa308404280fa0cc0211a6f9fb/LICENSE) — Checked on 2026-10-25 | | Checkpoint — [ntinosbarmpas/NeuroRVQ ↗](https://huggingface.co/ntinosbarmpas/NeuroRVQ/tree/d944b87f44ae0ba2923b2f10d0518f23f6803b76) at `d944b87f44ae0ba2923b2f10d0518f23f6803b76`: `pretrained_models/foundation_models/NeuroRVQ_EEG_foundation_model_v1.pt`, SHA-256 `d14ef4eae2d4594d5f1f709226a3bb5f362afd84b68050d5da0fa0586aea886a` | [CC BY-NC 4.0 ↗](https://creativecommons.org/licenses/by-nc/4.0/) — the NeuroRVQ checkpoints in the Hugging Face repository at the pinned revision, under the same licence file as the code. No weights are distributed; the SHA-256 identifies the checkpoint used and grants nothing. | [LICENSE ↗](https://huggingface.co/ntinosbarmpas/NeuroRVQ/blob/d944b87f44ae0ba2923b2f10d0518f23f6803b76/LICENSE) — Checked on 2026-10-25 | **The model, as its authors publish it.** [arXiv 2510.13068 ↗](https://arxiv.org/abs/2510.13068). NeuroRVQ is used here frozen, and the transfer is supervised: the frozen backbone’s pooled values feed a ridge head fitted on the other folds’ labels. The authors’ affine inference wrapper fails a strict restore: eight normalisation tensors it expects are missing from the checkpoint. What ran is the released backbone with parameter-free branch normalisation and mean-token pooling. ## EEGMAT in ten-second windows: two frozen encoders EEG-Captioner’s encoder and AXON, each kept frozen and read by one fixed ridge head on ten-second windows of EEGMAT’s arithmetic and rest, published as aggregate scores in a panel of its own on the pilot-studies page, outside the pilot count and every leaderboard. No new dataset: EEGMAT’s released record is reused unchanged, with what this release publishes on it in the table below. EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. EEG-Captioner’s source code is MIT. EEG-Captioner’s encoder weights are Apache-2.0 as its Hugging Face weight metadata states. AXON’s checkpoint is Apache-2.0 as its Hugging Face licence tag states; its Braindecode module carries per-file Apache-2.0 terms within a BSD-3-Clause project. No weights are distributed, and no endorsement by either model’s authors is implied. Pretraining exposure is unknown for both encoders. Not published: per-person, per-fold and per-window values of any kind: per-fold balanced accuracy, fold assignments, participant identifiers, features, fitted heads, predictions and the sealed files. Only participant-mean balanced accuracy, paired differences with their intervals, cohort counts and the source-QC totals are published; confusion matrices and pooled macro-F1: window-pooled summaries, a different aggregation from the participant means shown; the source-QC breakdown beyond the totals: the count of flagged channel-record instances and the counts by condition and by fold; workflow accounting: forward counts, head-fit counts, CPU and wall time and memory, from the handoffs, ledgers and audits; a difference between the two encoders: none is computed; a score on a cohort that meets EEG-Captioner’s source QC: no such study was selected, and one would need a new prospective protocol; AXON’s paper reference: not cited until checked; checkpoints and weights: none is distributed; each SHA-256 identifies the file used and grants nothing; the handoffs, protocols, approvals, reviews, audits and receipts themselves: each is pinned by SHA-256 in the release manifest and checked by the export, and none is published, because several carry private storage paths. Release [`eegmat-ten-second-update-20261028`](https://bci.report/releases/#eegmat-ten-second-update-20261028). | Dataset | Source terms | What is published | | --- | --- | --- | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — PhysioNet EEG During Mental Arithmetic Tasks 1.0.0 · 1.0.0 — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) — Personal noncommercial research; aggregate results only | **Aggregate scores** — Reused unchanged from the review of the same recordings for the core matrix. Published here for two frozen encoders on ten-second windows: participant-mean balanced accuracy of each encoder and of Welch band power, two paired differences with their participant-bootstrap intervals, cohort counts and the source-QC totals; no per-person, per-fold or per-window value. | The two encoders’ terms, part by part: | Part | Terms | As stated by | | --- | --- | --- | | EEG-Captioner code — [lyJinAoki/EEG-Captioner ↗](https://github.com/lyJinAoki/EEG-Captioner) at `088d9af2dedcb181534f7fdcff7c913a19e10f04` | MIT | EEG-Captioner’s source code at the pinned commit | | EEG-Captioner encoder weights — [LyJinAoki/EEG-Captioner ↗](https://huggingface.co/LyJinAoki/EEG-Captioner) at `1d56a371d68a108dd56c362cc41205c5b511cc76`, SHA-256 `060a5c0d9ac11543470038adbcbc5d9ba5828c1dc7c7d5097549435fbe28e989` | [Apache-2.0 ↗](https://www.apache.org/licenses/LICENSE-2.0) | the portable weight metadata of the Hugging Face repository at the pinned revision | | AXON checkpoint — [MannasAI/axon-eeg ↗](https://huggingface.co/MannasAI/axon-eeg) at `fe7f85239e6018f393039a10a103bf7075c93aa6`: `model.safetensors`, SHA-256 `89c864a3ed947588ffc740cd4b342a2158661b13898c0dd9a86abebd33ffc882` | [Apache-2.0 ↗](https://www.apache.org/licenses/LICENSE-2.0) — Not gated | the licence tag of the Hugging Face repository at the pinned revision | | AXON module — [braindecode/braindecode ↗](https://github.com/braindecode/braindecode) at `8c618f15a702f63d4899e3e8bc2e6f10a5f54151`: `braindecode.models.AXON` | Apache-2.0 | the per-file terms of the AXON module in Braindecode | | Braindecode project | BSD-3-Clause | the licence at the Braindecode project root | **Weights.** No checkpoint or weights are distributed: each SHA-256 identifies the file used and grants nothing. ## Dataset checks: Kim beta-range, MIND, Damm and NeBULA Four public EEG datasets on the data-quality page, each read from its own files, outside the core matrix and every leaderboard. Kim beta-range SSVEP is published as aggregate scores: training-free CCA on all its participants, with a dataset page. MIND is published as aggregate scores on record IDs, as a conditional case, with the licence statements of its sources listed in the table below. Damm finger motor imagery and NeBULA have one participant each and are published as design, counts and file facts; their scores are held, because with one person every score is that person’s. The licences of Kim beta-range, Damm and NeBULA were read from each source’s own record, and the export records each read with its capture’s SHA-256; for MIND, the table says how each statement was read. Each ethics and consent statement comes from the document named beside it. Not published: values per person, per record ID, per fold, per run or per trial: identifiers, file and run names, fold assignments, features, fitted models, decision values and predictions; raw signals from any of the four datasets; every Damm and NeBULA score: accuracies, balanced accuracies, control values, differences, paired correctness counts and diagnostic fractions, held by the owner’s decision under the one-person rule; and any comparison of NeBULA’s order diagnostic with its EEG or EMG arms; MIND’s pooled metrics, confusion matrices and macro-F1, the secondary panel’s schedule census and differences, the overfull runs’ marker counts and the model-fit counts; Kim’s per-participant accuracies, the fixture identities of the independent audit and the stored-value ranges of the schema record; compute times, CPU and memory accounting, storage sizes, hosts and storage locations; the research aggregates, audits, reviews, protocols and handoffs themselves: each is pinned by SHA-256 in the release manifest and checked by the export, never published, because several carry private storage paths. Release [`dataset-checks-update-20261028`](https://bci.report/releases/#dataset-checks-update-20261028). | Dataset | Source terms | What is published | | --- | --- | --- | | [Kim beta-range SSVEP: 40 targets from 14.0 to 21.8 Hz (Kim, Won, Ahn and Jun, Scientific Data) ↗](https://doi.org/10.6084/m9.figshare.28806815.v3) — Figshare record 28806815, version 3, the latest of the 3 versions the record lists — Heegyu Kim · SSVEP 40class dataset, figshare, doi:10.6084/m9.figshare.28806815.v3, CC BY 4.0. Paper: Heegyu Kim, Kyungho Won, Minkyu Ahn and Sung Chan Jun, A 40-Class SSVEP Speller Dataset: Beta Range Stimulation for Low-Fatigue BCI Applications, Scientific Data 12:1751 (2025), doi:10.1038/s41597-025-06032-2. Modifications: the 40 task files only; the mean of M1 and M2 subtracted, 11 channels, a 2-second window and a band-pass filter inside it; no data redistributed. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — As the licence field of Figshare record 10.6084/m9.figshare.28806815.v3 states it. The paper’s Data availability links version 2; the research used version 3 (the latest), which adds a questionnaire. Both versions carry CC BY 4.0 in the licence field. The paper’s own licence, Creative Commons Attribution 4.0 International License, covers the data descriptor article; the data licence is the Figshare record’s. | **Aggregate scores** — Ethics: Institutional Review Board at Gwangju Institute of Science and Technology (GIST), approval 20211201-HR-64-02-04. Consent: signed consent forms, including consent for the use and open sharing of the data. The confidentiality clause of the consent forms excludes questionnaire-related personal data; the research acquired the 40 task files only and excluded the questionnaire and rest files. Published here: the mean participant accuracy with its participant-bootstrap interval and aggregate counts; no per-participant value, identifier, prediction or signal. | | [MIND: A Multimodal fNIRS-EEG Dataset for Unilateral Limb Motor Imagery (ScienceDB 10.57760/sciencedb.34326) ↗](https://doi.org/10.57760/sciencedb.34326) — ScienceDB record 10.57760/sciencedb.34326, record version 4.0.0 — Lufeng Feng, Baomin Xu, Haoran Zhang, Bihai Lin, Zuxuan Deng, Sidi Tao, Chenyu Liu, Shifan Jia, Li Duan and Ziyu Jia · MIND: A Multimodal fNIRS-EEG Dataset for Unilateral Limb Motor Imagery, Science Data Bank, doi:10.57760/sciencedb.34326. Paper: arXiv 2602.04299. Modifications: EEG only, filtered over each whole run, resampled and cut into stored-marker windows; no data redistributed. Derived analysis by BCI Report; not endorsed by the authors. | CC BY 4.0 — the paper (quoted from the research intake receipt of the paper page); CC BY 4.0 — the archive dataset description (as the research protocol records the dataset description inside the downloaded archive); CC BY-NC 4.0 — the MOABB adapter (the licence field of the MOABB dataset adapter); [CC BY-NC 4.0 ↗](https://creativecommons.org/licenses/by-nc/4.0/) — the ScienceDB record page, read on 2026-10-28 (the schema.org metadata the record page serves) — These statements are set side by side, not reconciled. ScienceDB’s terms-of-use page was not read. Aggregate values only; no data are redistributed. | **Aggregate scores, conditional** — The paper states that the study protocol was approved by an ethics committee it does not name, with no approval number, and that participants signed written informed consent. Published here: record-ID mean balanced accuracies, record-ID resampling ranges and aggregate counts; no per-record-ID value, identifier, prediction or signal. | | [Damm finger motor imagery: Random and Sequential Order Finger Motor Imagery (OpenNeuro ds008446) ↗](https://doi.org/10.18112/openneuro.ds008446.v1.0.1) — OpenNeuro ds008446 at its latest snapshot, read at a pinned commit of its mirror — Laura Marie Damm, Dai Jiang and Andreas Demosthenous · Random and Sequential Order Finger Motor Imagery, OpenNeuro ds008446, doi:10.18112/openneuro.ds008446.v1.0.1, CC0. Funding: Engineering and Physical Sciences Research Council (EPSRC) grant EP/S022139/1. Modifications: the first participant’s two random-order runs only; no score from these data is published. Derived analysis by BCI Report; not endorsed by the authors. | CC0 — As the License field of [dataset_description.json at the pinned commit ↗](https://raw.githubusercontent.com/OpenNeuroDatasets/ds008446/8d8448e2dcd731b251f2f3c85b98cce5847cb000/dataset_description.json) states it. The README cites a LICENSE file that the tree at the pinned commit does not hold; the only licence statement is CC0 in dataset_description.json. | **Held — no figure** — Ethics: UCL Humanities, Arts and Sciences Research Ethics Committee, approval 26907/001. No consent statement was found in the dataset files or its metadata, and no paper is named. Published here: design, counts and file facts for one participant’s recordings and no score; no identifier, per-trial value or signal. | | [NeBULA (Neuromechanical Biomarkers for Upper Limb Assessment): an EEG-EMG dataset from a standardized reaching task (Garro et al., Scientific Data 12, 831) ↗](https://doi.org/10.6084/m9.figshare.27301629.v1) — Figshare record 27301629, version 1, the only version the record lists — Florencia Garro, Elena Fenoglio, Indya Ceroni, Inna Forsiuk, Michele Canepa, Michael Mozzon, Agnese Bruschi, Francesco Zippo, Matteo Laffranchi, Lorenzo De Michieli, Stefano Buccelli, Michela Chiappalone and Marianna Semprini · An EEG-EMG Dataset from a Standardized Reaching Task for Biomarker Research in Upper Limb Assessment, figshare, doi:10.6084/m9.figshare.27301629.v1, CC BY 4.0. Paper: Florencia Garro, Elena Fenoglio, Indya Ceroni, Inna Forsiuk, Michele Canepa, Michael Mozzon, Agnese Bruschi, Francesco Zippo, Matteo Laffranchi, Lorenzo De Michieli, Stefano Buccelli, Michela Chiappalone and Marianna Semprini, An EEG-EMG dataset from a standardized reaching task for biomarker research in upper limb assessment, Scientific Data 12:831 (2025), doi:10.1038/s41597-025-05042-4. Modifications: the first participant’s files only; no score from these data is published. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — As the licence field of Figshare record 10.6084/m9.figshare.27301629.v1 states it. The paper’s own licence, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, covers the data descriptor article only; the data licence is the Figshare record’s. The files’ own dataset description names Zenodo 13627307, which is no registered public record: the Zenodo API, the record page, DataCite and doi.org each return 404. The paper names Figshare alone. | **Held — no figure** — Ethics: local ethical committee CER Liguria, approval 784/2022 (database id 12044). Consent: informed consent before participation, including consent to share the data openly. Published here: design, counts and file facts for one participant’s recordings and no score; no identifier, per-trial value or signal. | **Ethics and consent, as each source states them.** Kim beta-range SSVEP — ethics approval: “The Institutional Review Board at Gwangju Institute of Science and Technology (GIST) approved this experiment (20211201-HR-64-02-04)” Informed consent: “all subjects were informed of all experimental procedures and signed consent forms” Read from [www.nature.com ↗](https://www.nature.com/articles/s41597-025-06032-2). MIND — ethics approval: “The study protocol was approved by the Ethics Committee. All procedures followed institutional review board (IRB) guidelines and relevant privacy regulations.” Informed consent: “After confirming that they understood the instructions, all participants voluntarily signed written informed consent forms.” Quoted from the research intake receipt of the paper page, [arxiv.org ↗](https://arxiv.org/html/2602.04299v1). The committee is not named and no approval number is given in the paper. Damm finger motor imagery — ethics approval: “UCL Humanities, Arts and Sciences Research Ethics Committee; Approval Number: 26907/001” Informed consent: not stated. No consent statement in README.md, dataset_description.json, CHANGES or the OpenNeuro dataset metadata at the pinned commit; neither the README nor the metadata names a paper, so no paper was read. Read from [raw.githubusercontent.com ↗](https://raw.githubusercontent.com/OpenNeuroDatasets/ds008446/8d8448e2dcd731b251f2f3c85b98cce5847cb000/dataset_description.json). NeBULA (Neuromechanical Biomarkers for Upper Limb Assessment) — ethics approval: “The study’s experimental procedures were approved by the local ethical committee (CER Liguria:784/2022 - DB id 12044).” Informed consent: “All participants gave consent to participate and openly share the data collected.” Read from [www.nature.com ↗](https://www.nature.com/articles/s41597-025-05042-4). ## Core matrix: classical, training-free and scratch rows Rows run under the core matrix’s exact protocols, published as aggregate scores in a dated file of their own, [core-matrix-20261106.json](https://bci.report/data/core-matrix-20261106.json), with one CSV per protocol: FBCCA on both BETA protocols; TS+LR, MDM, xDAWN+LDA and XDAWNCov+TS+LR on their core protocols and on idle; ShallowFBCSPNet, Deep4Net and CSP+LDA in cells the matrix had left blank. Release [`core-matrix-update-20261106`](https://bci.report/releases/#core-matrix-update-20261106). The seven core datasets keep their released records and terms; nothing was downloaded. pyRiemann 0.12, MNE, scikit-learn, braindecode, SciPy and NumPy are BSD-3-Clause or BSD-family; no weights are distributed. Trained network states stay private, and no code is distributed. Each protocol keeps its released dataset record, licence and attribution; nothing was downloaded for these rows. | Dataset | Protocol and the rows this file adds | Licence, as released | | --- | --- | --- | | ds003810 | [Motor imagery & rest](https://bci.report/protocols/mi-rest/#core-rows): TS+LR, MDM, ShallowFBCSPNet, Deep4Net | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | | ds005342 | [Idle & command](https://bci.report/protocols/idle/#core-rows): TS+LR, MDM | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | | BETA | [SSVEP · 8 channels](https://bci.report/protocols/beta-8ch/#core-rows): FBCCA, ShallowFBCSPNet, Deep4Net [SSVEP · 4 channels](https://bci.report/protocols/beta-4ch/#core-rows): FBCCA, ShallowFBCSPNet, Deep4Net | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | | EEGMAT | [Arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#core-rows): TS+LR, MDM, ShallowFBCSPNet, Deep4Net, CSP+LDA | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | | ds006593 | [P300 target ERP](https://bci.report/protocols/p300-target/#core-rows): XDAWNCov+TS+LR, xDAWN+LDA, ShallowFBCSPNet, Deep4Net, CSP+LDA | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | | TMNRED / ds005383 | [Semantic target ERP](https://bci.report/protocols/semantic-target/#core-rows): XDAWNCov+TS+LR, xDAWN+LDA, ShallowFBCSPNet, Deep4Net, CSP+LDA | [OpenNeuro metadata says CC0; accompanying publication/GitHub says CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) | | EESM19 scalp subset | [Sleep staging](https://bci.report/protocols/sleep-scalp/#core-rows): ShallowFBCSPNet, Deep4Net, CSP+LDA | [CC0-1.0 declared by upstream and mirror ↗](https://creativecommons.org/publicdomain/zero/1.0/) | **Software.** pyRiemann 0.12, BSD-3-Clause; braindecode 1.5.1 and 1.2.0, BSD-3-Clause; MNE, scikit-learn, SciPy and NumPy, BSD-family; MOABB reference pipelines, BSD-3-Clause, read for citation and not executed; FBCCA sub-band design after TRCA-SSVEP src/filterbank.m, MIT; the code that ran is this project's own SciPy realisation. pyRiemann 0.12 ran from the wheel with SHA-256 `74a90b2ef0938c94fe9bf90b6506991fc19da55850aae5fdf992026b3845d43a`. **FBCCA code credit.** FBCCA: Chen, Wang, Gao, Jung and Gao, J. Neural Eng. 12(4):046008, 2015; sub-band edges, orders and ripple follow Masaki Nakanishi’s TRCA-SSVEP src/filterbank.m (MIT, revision 35c1b8a7ebfa3774dd7266931784d04ae09f3461), with the MIT notice kept beside the code; the implementation is this project’s own SciPy realisation. **Idle rows.** In these rows the idle protocol’s cohort size, 4 people, prints as a protocol fact, and their count of people who always abstained stays empty. Left empty: counts of people meeting a criterion print only for cohorts of 10 or more. **Contrasts against an EEGNet cell** carry the label “against the matrix's one-seed EEGNet cell; training variance not included” once beside their numbers. **Not published.** values per person, per participant identifier, per fold, per trial or per window: predictions, decision scores, correlations, probabilities, confusion matrices, idle thresholds and idle counts by person; per-sub-band or per-frequency values; fitted xDAWN filters, prototypes, covariance reference points and logistic-regression or LDA coefficients; trained network states, CSP filters and LDA weights; counts of people meeting a criterion: the idle rows print the cohort size of four as a protocol fact and leave the abstain count empty; per-seed values and seed ranges: each network cell is one training run with the matrix seed; the idle mean test-window AUROC; run facts: timings, memory, condition numbers, machine and environment details; the seconds field is empty for every row of this file; bridge reproductions as results, and synthetic smoke-test numbers; the Bonferroni-width intervals of the seven ShallowFBCSPNet − Deep4Net contrasts, which are descriptive and outside the family; sensitivity arms as matrix rows; contrasts outside the reviewed sets; results outside these designs (other windows, other electrode sets, duration groups); private storage locations and machine names. ## Log band-power ridge beside the Spectral ridge row BCI Report’s own comparator, a log band-power ridge (5 bands, visible channels), published as aggregate scores beside the Spectral ridge row of seven core protocols: one section on each of those protocol pages and [a section on the methods page](https://bci.report/methods/#log-band-power-ridge), outside the core matrix, the home matrix, every leaderboard and the register. No new dataset: each protocol’s released record is reused unchanged, in the table below. The seven protocols keep their released data records (ds003810, ds006593 and the EESM19 scalp subset CC0-1.0; EEGMAT ODC-By 1.0; TMNRED CC BY 4.0, the stricter of its two stated terms; BETA CC BY 4.0). The comparison rows are cited by their published values under their published credits. The comparator is the project’s own code, run with numpy, scipy and scikit-learn (BSD-3-Clause); no third-party weights or code were run. Not published: per-window predictions, per-person scores, per-fold values and row identifiers; features, fitted scalers and heads, the ridge alpha chosen in each fold, and the inner validation scores; the values of two internal checks, a train-selected alpha and a microvolt input view, and their per-protocol differences: the methods entry prints their two approved summary lines; an input view without window-mean removal, which was never scored; bridge and validity-gate receipts beyond pass or fail; compute, hosts, paths and run times; the research protocol, aggregate, audit, reviews and handoffs themselves, pinned by SHA-256 in the release manifest. Release [`logpower-ridge-update-20261106`](https://bci.report/releases/#logpower-ridge-update-20261106). | Dataset | Source terms | Released record, reused unchanged | | --- | --- | --- | | ds003810 | CC0-1.0 | [`mi-rest-protocol.json`](https://bci.report/data/mi-rest-protocol.json) | | EEGMAT | ODC-By 1.0 | [`arithmetic-rest-protocol.json`](https://bci.report/data/arithmetic-rest-protocol.json) | | ds006593 | CC0-1.0 | [`p300-target-protocol.json`](https://bci.report/data/p300-target-protocol.json) | | TMNRED / ds005383 | CC BY 4.0, the stricter of the CC0 metadata and the CC BY 4.0 publication terms | [`semantic-target-protocol.json`](https://bci.report/data/semantic-target-protocol.json) | | EESM19 scalp subset | CC0-1.0 | [`sleep-scalp-protocol.json`](https://bci.report/data/sleep-scalp-protocol.json) | | BETA | CC BY 4.0 | [`beta-8ch-protocol.json`](https://bci.report/data/beta-8ch-protocol.json) · [`beta-4ch-protocol.json`](https://bci.report/data/beta-4ch-protocol.json) | ## Signal-JEPA and EEGSimpleConv beside the matrix Signal-JEPA, kept frozen and read by a ridge head fitted on the training people, and EEGSimpleConv, trained from scratch with the mean of three seeds published, as aggregate scores beside the core matrix on the [motor imagery & rest](https://bci.report/protocols/mi-rest/#signal-jepa-eegsimpleconv) and [arithmetic & rest](https://bci.report/protocols/arithmetic-rest/#signal-jepa-eegsimpleconv) protocol pages, each compared with EEGNet on the same people. No new dataset: each protocol’s released record is reused unchanged, in the table below. ds003810 (CC0-1.0) and EEGMAT (Open Data Commons Attribution License 1.0) keep their released records; aggregate results only. Signal-JEPA’s weights are MIT as its model card at revision 51232ee0 states; its code is BSD-3-Clause (braindecode 1.5.1); the weights are not redistributed. EEGSimpleConv’s code is BSD-3-Clause (braindecode 1.5.1) and the separate author repository’s licence is unclear; it was trained from scratch, and no trained weights are shared. The model directory’s notes say Signal-JEPA was not downloaded or executed, and EEGSimpleConv not executed, for that release. Both were run the same day in a separate study on motor imagery & rest and arithmetic & rest; their results are on those two protocol pages. The released file keeps its bytes, and the correction is listed on [the releases page](https://bci.report/releases/#corrections). Not published: per-person, per-fold and per-window values, predictions, features, fitted heads, trained weights and participant identifiers; EEGSimpleConv’s per-seed balanced accuracies, their spread and range, and its per-seed intervals: only the mean of three seeds is published; an interval for EEGSimpleConv’s difference: the earlier study computed intervals per seed only; measured run times, and the machines and storage paths of the runs; a partial fine-tuning arm, which was not run; the research aggregates, audits, reviews and handoffs themselves, pinned by SHA-256 in the release manifest because some carry private storage paths. Release [`signal-jepa-eegsimpleconv-update-20261106`](https://bci.report/releases/#signal-jepa-eegsimpleconv-update-20261106). | Dataset or part | Terms | Here | | --- | --- | --- | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) | Aggregate scores; released record reused | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) | Aggregate scores; released record reused | | Signal-JEPA weights | MIT, as the model card at revision 51232ee0 states | Not redistributed | | Signal-JEPA code | BSD-3-Clause (braindecode 1.5.1) | Not redistributed | | Signal-JEPA pretraining data | Lee2019 (OpenBMI), as its model card names it; its terms were not checked | Not used here | | EEGSimpleConv code | BSD-3-Clause (braindecode 1.5.1); the separate author repository’s licence is unclear | Not redistributed | | EEGSimpleConv weights | none; trained from scratch | None shared | ## LaBraM at Scale and REVE Small The frozen-probe cells of Quark-B, Quark-L, Quark-H, Quark-G, Atom-L and Molecule-L and REVE Small reuse the eight core protocols’ reviewed records unchanged; no other dataset is involved. The datasets keep their released terms: ds003810 CC0-1.0; ds005342 CC0-1.0; BETA CC BY 4.0; EEGMAT ODC-By 1.0; ds006593 CC0-1.0; ds005383 CC BY 4.0, the stricter of the two stated terms; EESM19 CC0-1.0. Only aggregates are published ([foundation-checkpoints-update.json](https://bci.report/data/foundation-checkpoints-update.json), release [`foundation-checkpoints-update-20261106`](https://bci.report/releases/#foundation-checkpoints-update-20261106)). **LaBraM at Scale.** Code and weights under MIT, ungated: weights `Pan-Junyu/LaBraM-at-Scale @ a0b0ce27` for Quark-B, Quark-L, Quark-H, Atom-L and Molecule-L, and for Quark-G (`quark/quark-giant.safetensors`, Local Ubuntu / CUDA); code `AlbertMentat/LaBraM-at-Scale @ faa387d9`. Cite [arXiv 2611.03165](https://arxiv.org/abs/2611.03165) and, for the architecture, LaBraM ([arXiv 2405.18765](https://arxiv.org/abs/2405.18765)). **REVE Small.** `reve-small @ 3565bdcb` under the REVE Responsible Use License v1.0, accepted by the owner as for REVE Large above, with its conditions: aggregate scientific results only; no re-identification, inversion, membership inference or profiling; cite the REVE paper and name the model version; no adapted weights. Cite [arXiv 2510.21585](https://arxiv.org/abs/2510.21585). LaBraM at Scale code and weights are MIT and ungated; REVE Small is under the REVE Responsible Use License v1.0, aggregate scientific results only; no weights are distributed. Not published: per-person, per-fold and per-window values of any kind; features, fitted heads, idle thresholds and participant identifiers; counts of people meeting a criterion on the idle protocol, whose cohort of four prints as a protocol fact beside the trial counts; contrasts that were not computed: Atom-L or Molecule-L against Quark, REVE Small against REVE Base or Large, and an average of Quark-H minus Quark-B; Quark-G against Quark-B or Quark-L, any sum of ladder steps, and a GPU restatement of the Quark-L and Quark-H steps, with anything derived from them; the GPU re-run of Quark-H as cells: it is the baseline of the Quark-G step, and the Quark-H cells shown are the CPU cells; the re-extraction and attention-kernel checks as results; they are a methods note; context values of the numeric checks: an unweighted mean over protocols, intervals from an unrelated bootstrap seed, a numeric order of the BETA participants and training-fold sizes; weights of any kind: none were adapted, and downloaded weights are never redistributed; timings, memory, device readings, free-space values, the compute estimate and the download time from the runs; the research records themselves, pinned by hash in the release manifest because several carry private storage paths. **Random-initialization controls** ([random-init-controls-update.json](https://bci.report/data/random-init-controls-update.json), release [`random-init-controls-update-20261106`](https://bci.report/releases/#random-init-controls-update-20261106)): no other dataset and no other model. The random initializations are built from each model’s own constructor and are never distributed. REVE Large under the REVE Responsible Use License v1.0 (reve-large @ 317531c7): aggregate scientific results only. ST-EEGFormer, SingLEM and EEGMamba under MIT. Downloaded and random weights stay local. Not published: per-seed values, seed ranges and seed standard deviations: only the mean of the three random initializations is published; per-person, per-fold and per-window values; features, predictions, fitted heads and participant identifiers; counts of people helped or harmed: none were computed; random or adapted weights of any encoder: random states stay local, and downloaded weights are never redistributed; the alpha the train-selected readout chose in each fold; timings, memory and device readings from the run; bridge and gate details beyond the methods notes; the research records themselves, pinned by SHA-256 in the release manifest because several carry private storage paths. ## NSP components: a separate pilot panel A local test of the components of Neural State Prediction (arXiv:2609.31167v1, cited only), in a panel of its own on the pilot-studies page, after the counted pilots: [the panel](https://bci.report/topics/pilot-studies/#nsp-components). No new dataset: EEGMAT’s and ds003810’s released records are reused unchanged, with what this release publishes on each in the table below. EEGMAT keeps its released record: Open Data Commons Attribution License 1.0, aggregate results only. ds003810 keeps its released record: CC0-1.0, aggregate results only. eeg-fm-masking’s weights are CC BY 4.0 (cite arXiv:2609.33487) and its code is MIT. The NSP paper, arXiv:2609.31167v1, is CC BY-NC-ND 4.0 and is cited only; nothing from it is reproduced. The local model follows the paper’s MIT-licensed reference code. No weights are distributed, and no endorsement by the authors of either paper is implied. Not published: per-person, per-fold, per-seed and per-window values of any kind: fold and seed-by-fold results, each job’s diagnostics and their spread, participant identifiers, the selected label rows of the fitting-label subsets, features, fitted heads, predictions, checkpoints and the sealed files. Only equal-person means, paired differences with their descriptive intervals, probe and cosine-drop means over jobs, and counts are published; any figure or text of the NSP paper, which is cited only; compute, hosts, storage, paths, run times and memory; checkpoints and weights: none is distributed; each SHA-256 identifies the file used and grants nothing. Release [`nsp-components-update-20261106`](https://bci.report/releases/#nsp-components-update-20261106). | Dataset | Source terms | What is published | | --- | --- | --- | | [EEGMAT ↗](https://physionet.org/content/eegmat/1.0.0/) — PhysioNet EEG During Mental Arithmetic Tasks 1.0.0 · 1.0.0 — Igor Zyma, Ivan Seleznov, Anton Popov, Mariia Chernykh, Oleksii Shpenkov · EEG During Mental Arithmetic Tasks 1.0.0, PhysioNet, doi:10.13026/C2JQ1P. Study: Zyma et al. (2019), doi:10.3390/data4010014. PhysioNet platform: Pollard et al. (2026), doi:10.1038/s44360-026-00096-z. | [Open Data Commons Attribution License 1.0 ↗](https://opendatacommons.org/licenses/by/1-0/) — Personal noncommercial research; aggregate results only | **Aggregate scores** — Reused unchanged from the released record of the same recordings. Published here for the NSP component panel: equal-person balanced accuracies, paired differences with their participant-bootstrap intervals, probe and cosine-drop means over jobs, and cohort counts; no per-person, per-fold, per-seed or per-window value. | | [ds003810 ↗](https://doi.org/10.18112/openneuro.ds003810.v2.0.2) — OpenNeuro snapshot 2.0.2 · c674303319303b9ffc4ff7e2340b3a4807e8ffb5 — Peterson et al. · OpenNeuro ds003810, version 2.0.2. Study: https://pmc.ncbi.nlm.nih.gov/articles/PMC9114495/ | [CC0-1.0 ↗](https://creativecommons.org/publicdomain/zero/1.0/) — Personal noncommercial research; aggregate results only | **Aggregate scores** — Reused unchanged from the released record of the same recordings. Published here for block 3 of the NSP component panel: equal-person balanced accuracies, paired differences with their participant-bootstrap intervals, and cohort counts; no per-person, per-fold or per-window value. | The models’ and the paper’s terms: | Part | Terms | Basis | | --- | --- | --- | | eeg-fm-masking weights (MAE and JEPA, r = 9 cm, L = 2) | CC-BY-4.0 | [cite arXiv:2609.33487 ↗](https://arxiv.org/abs/2609.33487) | | eeg-fm-masking code | MIT | [Repository ↗](https://github.com/PierreGtch/eeg-fm-masking) — at `38a089e63d019c23793201fff58a27d7ce399c1a` | | The NSP paper, arXiv:2609.31167v1 | [CC BY-NC-ND 4.0 ↗](https://creativecommons.org/licenses/by-nc-nd/4.0/) | [arXiv record ↗](https://arxiv.org/abs/2609.31167v1) — Cited only; nothing from it is reproduced. | | The paper’s reference code | MIT (reference code, not linked) | As the paper’s reference code states; not linked | No checkpoint or weights are distributed; each SHA-256 identifies the file used and grants nothing. ## Fewer electrodes: a simulated cap shift on BETA A simulation on one public dataset, published as aggregate scores in its own section of the [fewer-electrodes page](https://bci.report/topics/fewer-electrodes/#cap-shift). It adds no dataset and no model: BETA keeps its released rights record, and the six decoders are the ones already described on their method pages. Not published: per-person, per-fold, per-window and per-seed values, including EEGNet’s per-seed means and any spread across seeds; participant identifiers, and which channel was flagged for whom; predictions, scores, fitted models and the sealed files; counts of people for a 5-point change, and medians; intervals for the analysis without the three flagged people; the 1-s spline results; accuracies of the interpolated (spline) arms and of the shifted 1-s arms as levels; they enter only through contrasts: minus the original cap, minus the neighbour map at the same angle (the spline steps), or deployed minus recalibrated (1-s eTRCA); storage locations, hosts, compute times and the research records themselves (pinned by SHA-256 in the manifest). Release [`electrode-shift-update-20261106`](https://bci.report/releases/#electrode-shift-update-20261106), data [`electrode-shift-update.json`](https://bci.report/data/electrode-shift-update.json). | Dataset | Source terms | What is published | | --- | --- | --- | | [BETA ↗](https://huggingface.co/datasets/Bingchuan/BETA) — Bingchuan Liu et al. · BETA: A Large Benchmark Database Toward SSVEP-BCI Application (2020), doi:10.3389/fnins.2020.00627. Figshare 12264401 v3; mirror Bingchuan/BETA. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — Participant consent described non-commercial scientific research. | **Aggregate scores** — BETA keeps its released record: CC BY 4.0, cohort aggregates only, with participant consent described for noncommercial scientific research. CBraMod: code MIT, weights Apache-2.0 as its official model card states; no weights are distributed. Ensemble TRCA is credited to Nakanishi et al. 2018 and filter-bank CCA to Chen et al. 2015. | **CBraMod in this section.** Frozen, as on its method page: code MIT, weights Apache-2.0 as its official model card states; no weights are distributed. Ensemble TRCA is credited to Nakanishi et al. 2018 and filter-bank CCA to Chen et al. 2015. **The small-cohort note.** The electrode-shift section counts people losing or gaining 10 points or more out of 70, as a description, and names no one; three people flagged by a label-free channel check are described only as a group. ## Data quality: eldBETA, all participants The eldBETA record on the data-quality page, read for all 100 of its participants and published as cohort aggregates: training-free CCA and filter-bank CCA, with their file checks. The record keeps the licence it was read under; the page prints no participant-level value and no signal. Not published: per-participant accuracies and correct counts, and any ordering of participants; participant identifiers of scored or excluded people; the earlier one-participant score, which stays unrestated; per-trial predictions, correlations, sub-band values, per-block or per-trial file-check values, windows and signals; age, gender, BCIQ and SNR, which were never read; per-target accuracies, confusions and ceiling counts beyond the equal group; compute times, storage sizes, hosts and locations; the research records, pinned by SHA-256; no eldBETA data are redistributed. Release [`eldbeta-cohort-update-20261106`](https://bci.report/releases/#eldbeta-cohort-update-20261106). | Dataset | Source terms | What is published | | --- | --- | --- | | [eldBETA: A Large Eldercare-oriented Benchmark Database of SSVEP-BCI for the Aging Population (Liu, Wang, Gao and Chen) ↗](https://doi.org/10.6084/m9.figshare.18032669.v3) — Figshare record 18032669, version 3, the latest of the 3 versions the record lists | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — As the licence field of the Figshare record states it, version 3 of 3, the latest; the capture’s SHA-256 is recorded in the export. The data descriptor’s own licence, Creative Commons Attribution 4.0 International License, covers the data descriptor article; the data licence is the Figshare record’s. | **Cohort aggregates** — The protocol followed the Declaration of Helsinki and was approved by the institutional review board of Tsinghua University, approval number 20210032. The participants were familiarised with the protocol and gave full written consent before the experiment. Published here: file-check counts for all 100 participants’ recordings and cohort means with their intervals; no identifier, per-participant or per-trial value, or signal. | **Ethics and consent, as the data descriptor states them.** Ethics approval: “The experimental protocol was under the declaration of Helsinki and approved by the institutional review board of Tsinghua University (NO. 20210032).” Informed consent: “The participants were instructed to be familiarized with experimental protocol and gave full written consent before the experiment.” Read from [nature.com ↗](https://www.nature.com/articles/s41597-022-01372-9). The descriptor also states: “The eldBETA database is open-access for research and can be downloaded from the website” (data descriptor, abstract) **Modifications made here.** decoding of the original recording format with the header calibration; fixed channel and window selection; window-local filter-bank filtering for FBCCA; cohort-level accuracies published, no participant-level values. Filter-bank CCA is credited to Chen et al. 2015, doi:10.1088/1741-2560/12/4/046008; its filter design follows Nakanishi’s MIT-licensed implementation; notice kept. No code is distributed. **The converter code.** The converter behind the descriptor’s MATLAB epochs, Figshare 17260877, version 6 (CC BY 4.0), was read for lineage only; never executed or published. ## Clinical groups: ASEEG A schizophrenia case-control comparison on one public dataset, published as aggregate scores in its own section of the clinical-groups page beside age-and-sex comparators. ASEEG is a new source; it adds no dataset page and no count. The clinical bar above applies to it: Research results on a public data set. Not a diagnosis, not diagnostic accuracy, and not a medical device or medical advice. Not published: participant rows of any kind: file codes, ages, sexes, diagnoses, fold assignments, features, out-of-fold probabilities and fitted heads; per-fold and per-split-seed values, the split seeds themselves, and seed ranges; per-group demographics: age means, medians and ranges, sex counts, and any figure derived from them, the data descriptor’s own included; confusion counts: seed-averaged values are not whole people; recording lengths of individual files; bridge numbers beyond pass or fail and their largest differences, and smoke-test numbers; EEG signals, the descriptor’s figures or text beyond short quotes, and model weights; runtimes, hosts and storage locations; the research records themselves are pinned by SHA-256 in the release manifest. Release [`clinical-aseeg-update-20261106`](https://bci.report/releases/#clinical-aseeg-update-20261106). | Dataset | Source terms | What is published | | --- | --- | --- | | [ASEEG · Atieh Schizophrenia EEG ↗](https://doi.org/10.5281/zenodo.18029536) — Zenodo 18029536, First version, 2025-12-23, record revision 6 — Bagherzadeh S, Norouzi M, Farokhshad F, Tolou Koroushi P, Ghasri A, Bahri Hampa S, Kazemi R, Rostami R, Shalbaf A. ASEEG: Atieh Schizophrenia EEG. Zenodo, doi:10.5281/zenodo.18029536 (First version, 2025-12-23), CC BY 4.0. Data descriptor: Scientific Data 13:1211 (2026), doi:10.1038/s41597-026-07968-9. Modifications: the first 256 s of each released recording read as released; log relative band power and frozen CBraMod features; cohort-level scores only. Derived analysis by BCI Report; not endorsed by the authors. | [CC BY 4.0 ↗](https://creativecommons.org/licenses/by/4.0/) — As stated by Zenodo metadata.license.id (cc-by-4.0) and the data descriptor, Data availability. The data descriptor is CC BY-NC-ND 4.0 (Scientific Data 13:1211 (2026)), so only short quotes appear on the site. | **Aggregate scores** — The data descriptor states that the ethics review board at Shahid Beheshti University of Medical Sciences approved the study, under ethics code IR.SBMU.MSP.REC.1404.441, and that specific informed consent covered the recording and the sharing of de-identified data, with written authorization from a parent, guardian or responsible adult where needed. The publication boundary is the one the Parkinson’s and IESSEEG sections use: aggregate scores only, with no participant rows or predictions and no per-group age or sex composition. | **Ethics and consent, as the data descriptor states them.** “All studies and activities are conducted under the ethics code number IR.SBMU.MSP.REC.1404.441 and adhere to the principles of ethical research.” “specific informed consent was obtained for the EEG recording procedure and for the subsequent sharing of de-identified data.” “a parent, legal guardian, or other directly responsible adult provided written authorization on a dedicated consent form.” Read from [the data descriptor ↗](https://doi.org/10.1038/s41597-026-07968-9); the ethics review board it names is Shahid Beheshti University of Medical Sciences. **Where the recordings come from.** “The participant group was sourced from the Atieh Clinic, a private facility located in Tehran, Iran.” “EEG signals were obtained retrospectively from the clinic’s medical records.” Both are the descriptor’s words. **CBraMod in this section.** CBraMod is read frozen, as a feature extractor, by the section’s one probe (Wang et al., arXiv 2412.07236); its code is MIT licensed and its weights Apache-2.0 (official model card), and no weights are distributed. The checkpoint is identified by its SHA-256 (`0792cb808c14e6b7a2bb2ce1dff379bc47bc54c49a779825bdfeb33bf8157178`). ## The open register The open register lists every EEG dataset we know of, and every EEG model and paper we track, with the furthest step checked on its ladder and what that step showed. Listing is open: a row can record only that a dataset exists, where it is hosted and what licence its host states, and the register also lists sources that give a reader nothing to run, such as L-FAME, because it records what was checked. The publication rules on this page are unchanged: a score enters a table on this site only after its source review, as an aggregate result, with nothing upstream redistributed. The two ladders are on [the dataset register](https://bci.report/datasets/register/) and [the model register](https://bci.report/methods/register/), and the rows are downloadable as [open-register-20261106.json](https://bci.report/data/open-register-20261106.json) and [open-register-20261106.csv](https://bci.report/data/open-register-20261106.csv), release [`open-register-update-20261106`](https://bci.report/releases/#open-register-update-20261106). **The register’s rules.** Gated data, or data that needs a signed or requested agreement, stays at step 1 unless we accept its terms. Suspected re-uploads stay at step 1 whatever their gating. They are compared by metadata only, and none of their files is downloaded. A finding about a re-upload is stated only when an independent check confirms it and a captured page of the original terms backs it. Rows state facts. Every row has a correction route. A row states what its host record, its files or its source’s own documents say, each with its source, and its step is the furthest one whose evidence is on file, with its gaps listed. Repository ids and usernames appear in a row’s text; a row never links to a repository or to an uploader’s page, and a Hugging Face row links its name to a Hub search for that name; no id or username becomes an anchor; an uploader’s series of repositories is one family row; card text, download counts, modification dates and demographics are not shown. **Correcting a row.** Every row links a mail to [contact@bci.report](mailto:contact@bci.report?subject=Register%20row) whose subject names the row’s opaque id and nothing else. Anyone can ask for a row to be corrected, or for a finding to be checked again, including the uploader or the authors. A corrected row appears in the next dated register file; the earlier file keeps its bytes and its hash on the releases page, as every release does. ## What stays local - Recordings, timestamps, subject mappings, labels and individual predictions remain outside the website. - We do not publish diagnosis-specific small groups, identifiable example traces, or demographic profiles. Aggregate scores describe decoder performance, not a person’s health or ability. - Source files with unclear provenance, incompatible terms or unresolved physical calibration do not enter published comparisons. Removing a name or replacing it with a code does not by itself anonymize EEG. Open licenses address specified intellectual-property rights; they do not generally clear privacy and personality rights. **An average over four people is close to its parts.** The idle protocol has a cohort of four, and at that size a published mean can be turned back into a count: 1.7% of 60 command trials is one detection, and since two of the four participants selected an always-abstain threshold, that detection belongs to one of the remaining two. We publish the abstention counts anyway, because a false-activation rate of zero produced by people who never activate is the more misleading number. What a reader cannot recover is which person is which — participants are unidentified here and in the public source recordings, and no demographic, session or ordering information is published that would let anyone line them up. **One animal, two folders, five and eight people.** The same arithmetic holds for the smallest cohorts beyond the core matrix. MC_RTT is one macaque: every figure is that animal’s own, and no interval is reported; it is animal data, with no participant to identify. REFED’s two nominal participant folders share the same videos, and each published error is a mean over all of their clips together: no per-folder or per-clip error is published, because with two folders one of them and the published mean would give away the other. The Orthosis pilot has eight people: its intervals resample whole people, how many of them had EEG above the confound probe is published as a count, and no per-person minimum, maximum or range is. BINGO’s subject fields, as the source defines them, are not proven distinct people, and only means over all of them, with descriptive intervals, are published. The handwriting analysis has five people: its intervals resample whole people, and each bound is a percentile of resampled five-person means, not one person’s score. Its mask-only control is the exception: its scores sit so close to chance that its bounds would show individual people’s scores, so that interval is not published. The status-only sources with one participant each, FALCON H2 and the functional-ultrasound recordings, publish counts of trials, files, runs and frames that describe the public files, not the person; so do the Damm and NeBULA cases on the data-quality page, with counts of cues, trials, channels and markers. **A median over an odd number of records is one record’s value.** The later-sessions page prints the cursor-tracking ridge’s median error beside its mean for each Forenzo cohort and response arm, and the Main cohort has an odd number of admitted records, so each of its printed medians is the error of one record. We publish those medians anyway, because a mean pulled far upwards by a few very large errors is the more misleading number without them. It does not identify anyone: the records are not proven to be distinct people, the page does not say which record the median is, and no participant, record or file identifier is published. ## How to interpret the comparisons Each protocol states the cohort, selected electrodes, preprocessing, split and training budget. Frozen foundation encoders, scratch-trained networks and classical references are labeled separately. The stable matrix keeps its fixed published seeds; the deployment extension separately reports five three-seed EEGNet sensitivity groups. These comparisons and descriptive intervals do not establish optimal performance or clinical validation. Pretraining exposure is stated as what each model’s authors publish as their pretraining list, checked on 4 October 2026 (EEGPT’s on 12 October 2026, LaBraM at Scale’s on 5 November 2026, from its paper’s corpus table) and linked to the source on the protocol and method pages: a dataset absent from the list is “not in the authors’ published pretraining list”, which is not proof that its recordings were never seen, and where the authors publish no list (ZUNA 1.1) it is unknown. Four- and eight-electrode subsets of laboratory recordings do not validate a physical four- or eight-channel headset. ## Corrections and changes of use Scientific corrections, benchmark-method questions, and rights, privacy or attribution concerns are all welcome, including ones that would withdraw a published result. Affected results can be withheld while an issue is reviewed. **Report an issue** Write to [contact@bci.report](mailto:contact@bci.report?subject=%5BBCI%20Report%20correction%5D%20) for scientific corrections and benchmark questions, or [privacy@bci.report](mailto:privacy@bci.report) for privacy, licensing and withdrawal requests. This domain receives mail but does not send it, so a reply will come from the maintainer’s own mailbox rather than from @bci.report. Replying to the address above still reaches the same place. When reporting, please include the page URL, the dataset and model, the release or protocol ID, a short description, and a public evidence link. **Do not send raw EEG, participant names, health records, passwords or identity documents**: an email address cannot refuse attachments. Identity proof is not requested. Published corrections are redacted scientific notes; the original message and sender details are not published. Advertising, sponsorship, paid access, new hosting jurisdictions, data uploads or clinical applications require a new scope review. A noncommercial source label is not a blanket authorization for those uses. ## Website privacy The application does not run analytics, advertising trackers or EEG processing. Hosting and access-control providers may process connection and sign-in information independently. External source links lead to sites with their own privacy practices. This site is not a medical service and provides no diagnosis, treatment advice or participant-level assessment. ## Useful references [Hugging Face terms](https://huggingface.co/terms-of-service) · [Creative Commons rights and attribution](https://creativecommons.org/licenses/by/4.0/legalcode.en) Core sources reviewed: 2026-09-20. --- Markdown copy of https://bci.report/data-use/, generated from the published page. Figures are aggregate results; terms of use: https://bci.report/data-use/