67.6%62.8%–72.5%Single seed — the highest of the seeds run (see Stability)
0.637Mean across held-out participants
19
36
117.7 s
TS+LRClassical method · OAS · 4–40 Hz · primary row · Local Ubuntu / CPU
Supervised fit
65.5%59.6%–71.8%
0.592Mean across held-out participants
19
36
—
MDMClassical method · OAS · 4–40 Hz · secondary row · Local Ubuntu / CPU
Supervised fit
71.5%65.5%–77.8%
0.661Mean across held-out participants
19
36
—
ShallowFBCSPNetCompact model · matrix recipe · AdamW · Apple M5 / MPS
Scratch · 20 epochs
67.9%63.4%–72.4%
0.644Mean across held-out participants
19
36
—
Deep4NetCompact model · matrix recipe · AdamW · Apple M5 / MPS
Scratch · 20 epochs
61.3%57.5%–65.2%
0.551Mean across held-out participants
19
36
—
CSP+LDAClassical method · broadband CSP · shrinkage LDA · Apple M5 / CPU
Supervised fit
68.3%62.0%–74.3%
0.649Mean 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, FBCCA (they need the stimulation frequencies, and this protocol has none); Temporal ridge, XDAWNCov+TS+LR, xDAWN+LDA (they model the waveform of a stimulus-locked transient potential, and this protocol's classes are not defined by one).
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.
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
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 ridgeTS+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.
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.
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.
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%)
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
53.3%49.5%–57.1%Interval reaches chance level
0.510
19
36
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 ↗
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.
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 ↗
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 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.
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.
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%)
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
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.
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.
its pretraining corpus includes this protocol's dataset
Molecule-LLocal 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 SmallLocal 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).
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.
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-JEPAFoundation 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
EEGSimpleConvCompact 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 ppvs 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 ↗ · model card ↗
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 ↗
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 ↗
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).
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.
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
CreditIgor 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.
Public-data register noteAttribution license and study-specific approval/consent are documented; eligibility assumes aggregate-only output and exclusion of subject-info fields.
Correction on record · : 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 →
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.