FBCCAClassical method · FBCCA-M3-N7-H5 · Local Ubuntu / CPU, training-free
Known-frequency reference · no fitting
79.1%75.1%–82.8%
0.789Mean across held-out participants
8
70
—
ShallowFBCSPNetCompact model · matrix recipe · AdamW · Apple M5 / MPS
Scratch · 20 epochs
23.0%20.3%–25.9%
0.200Mean across held-out participants
8
70
—
Deep4NetCompact model · matrix recipe · AdamW · Apple M5 / MPS
Scratch · 20 epochs
52.8%47.3%–58.2%
0.513Mean 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. 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.
Inspect point estimates and 95% intervals side by side on a shared axis. Paired differences appear only where a paired interval is published. Dropdowns are grouped by family alphabetically — never ordered by score.
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 →Fewer electrodes →
FBCCA: FBCCA-M3-N7-H5 · seven Chebyshev sub-bands, five harmonics · training-free
ShallowFBCSPNet and Deep4Net under the matrix recipe
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 CCAFBCCA-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 − EEGNetFBCCA-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.
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%.
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.
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%)
0%25%50%75%100%
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.
not in the authors' published pretraining list (checked 2026-10-04)source ↗
ZUNA 1.113Foundation modelModel 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 unknownunknown: the authors do not list their pretraining data closely enough to decidesource ↗
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.
not in the authors' published pretraining list (checked 2026-10-04)source ↗
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 ↗
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 ↗
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 ↗
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 ↗
SingLEM: MIT (checkpoint committed in the MIT repository)MIT.paper ↗
ZUNA 1.1: Apache-2.0; model card: research use only, not for diagnosis or clinical useApache-2.0 weights; model card states research use only, not for diagnosis or clinical use.paper ↗
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.
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.
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%)
0%25%50%75%100%
Balanced accuracy under this protocol. Grey: the released rows above; rust: EEGPT. Dot: the estimate; line: 95% interval; dashed line: chance level.
not in the authors' published pretraining list (checked 2026-10-12)source ↗
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.
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.
† 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).
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.
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. Pretraining exposure ↓
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.
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
CreditBingchuan 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.
Public-data register noteEligible only under the stated personal, research-led, noncommercial operation. Any ads, sponsorship, fees, or use directed toward commercial advantage requires fresh review or permission.
Hugging Face mirror of BETA database, Figshare record 12264401 v3 · mirror d4290c0200db8a104e0f557349dc49f90ba79506 · upstream Figshare version 3, 2022-06-15
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.