CBraMod: results on public EEG datasets

Foundation model

CBraMod: results on public EEG datasets

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; 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 · Protocols: Motor imagery & rest; Arithmetic & rest; SSVEP · 8 channels; SSVEP · 4 channels; P300 target ERP; Semantic target ERP; Sleep staging; Idle & command

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. ↗

Directory status: Evaluated · Official code ↗ · Description sources arxiv.org · github.com · huggingface.co

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 · Core matrix · motor imagery & rest

Source page: Core-matrix protocol: Motor imagery & rest · Chance level 50.0% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	Rest versus right-hand imagery	Balanced accuracy	62.9%59.6%–66.5%	10

ds003810 · Pretraining · fixed readout

Source page: Does pretraining help? · deployment-topics.json

Method	Condition	Metric	Value	People
CBraMod	MI / rest · pretrained	Balanced accuracy	62.9%	10
CBraMod	MI / rest · random initialization (mean)	Balanced accuracy	56.5%	10

ds003810 · Pretraining · train-selected readout

Source page: Does pretraining help? · deployment-topics.json

Method	Condition	Metric	Value	People
CBraMod	MI / rest · pretrained	Balanced accuracy	62.7%	10
CBraMod	MI / rest · random initialization (mean)	Balanced accuracy	54.9%	10

ds003810 · Reliable decisions · when to decline a decision (crude: ten people)

Source page: When not to act · reliable-decisions-update.json

Method	Condition	Metric	Value	People
CBraMod	Fixed threshold: calibrated confidence at least 0.8	Windows accepted	7 / 1,2006 of 10 people had nothing accepted.	10
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 ppNo 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). 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 →

EEGMAT · Core matrix · arithmetic & rest

Source page: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	Serial subtraction versus resting EEG	Balanced accuracy	62.3%58.1%–66.5%	36

EEGMAT · Pretraining · fixed readout

Source page: Does pretraining help? · deployment-topics.json

Method	Condition	Metric	Value	People
CBraMod	Mental workload · pretrained	Balanced accuracy	62.3%	36
CBraMod	Mental workload · random initialization (mean)	Balanced accuracy	56.4%	36

EEGMAT · Pretraining · train-selected readout

Source page: Does pretraining help? · deployment-topics.json

Method	Condition	Metric	Value	People
CBraMod	Mental workload · pretrained	Balanced accuracy	68.4%	36
CBraMod	Mental workload · random initialization (mean)	Balanced accuracy	61.7%	36

BETA · Core matrix · SSVEP, 8 channels

Source page: Core-matrix protocol: SSVEP · 8 channels · Chance level 2.5% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	40 visual targets · 8 posterior electrodes	Balanced accuracy	33.7%29.8%–37.6%	70

BETA · Core matrix · SSVEP, 4 channels

Source page: Core-matrix protocol: SSVEP · 4 channels · Chance level 2.5% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	40 visual targets · 4 posterior electrodes	Balanced accuracy	27.6%24.1%–31.3%	70

BETA · Reliable decisions · when to decline a decision

Source page: When not to act · reliable-decisions-update.json

Method	Condition	Metric	Value	People
CBraMod	Fixed threshold: calibrated confidence at least 0.8	Trials accepted	165 / 11,20038 of 70 people had nothing accepted.	70
CBraMod	Learned reject option minus calibrated confidence, same classifier	Error among the 80% most certain, difference	−0.03 pp−0.8 pp–+0.6 ppNo 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). 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 →

BETA · Questions in language · SSVEP (route 3)

Source page: Questions in language · questions-in-language-update.json

Method	Condition	Metric	Value	People
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 ppID higher; 2 pp margin not met; passes the triviality canary.	70
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 ppID higher; 2 pp margin not met; passes the triviality canary.	70
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 ppTPL(train) higher; 2 pp margin not met; passes the triviality canary.	70
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 ppTPL(train) higher; equivalent within ±2 pp; passes the triviality canary.	70
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 ppDESC(train) higher; 2 pp margin not met; passes the triviality canary.	70
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 ppUnseen question. CCA higher; 2 pp margin not met; headroom gate passed.	70
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 ppUnseen question. NN(TPL) higher; difference only: no margin; headroom gate passed.	70
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 ppUnseen question. NUM higher; difference only: no margin; headroom gate passed.	70
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 ppUnseen question. CCA higher; 2 pp margin not met; headroom gate passed.	70
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 ppUnseen 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 →

ds006593 · Core matrix · P300 target ERP

Source page: Core-matrix protocol: P300 target ERP · Chance level 50.0% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	Visual target versus nontarget events	Balanced accuracy	55.0%53.3%–56.7%	21

TMNRED / ds005383 · Core matrix · semantic target ERP

Source page: Core-matrix protocol: Semantic target ERP · Chance level 50.0% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	Reading · target versus nontarget events	Balanced accuracy	55.8%54.2%–57.4%	30

EESM19 scalp subset · Core matrix · sleep staging

Source page: Core-matrix protocol: Sleep staging · Chance level 20.0% · experiments.json

Method	Condition	Metric	Value	People
CBraMod	Five stages · balanced scalp-EEG sample	Balanced accuracy	71.3%68.7%–73.7%	20

ds005342 · Core matrix · idle & command

Source page: Core-matrix protocol: Idle & command · experiments.json

Method	Condition	Metric	Value	People
CBraMod	Seated motor imagery · cue-gated replay	Command detection ≤3 s	18.3%	4
CBraMod	Seated motor imagery · cue-gated replay	Idle false activation	1.7%	4

Mobile BCI dataset (SSVEP and ERP paradigms) · Movement · SSVEP, 2-second windows

Source page: On the move · deployment-topics.json

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%)

Balanced accuracy, 3 configurations. Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level.

Method	Condition	Metric	Value	People
CBraMod	Standing · scalp · 8 ch	Balanced accuracy	59.9%54.3%–65.4%	23
CBraMod	Slow walk · 0.8 m/s · scalp · 8 ch	Balanced accuracy	46.2%41.1%–51.4%	23
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) · One model, several questions · motor imagery (route 2)

Source page: One model, several questions · Chance level 50.0% · shared-representation-update.json

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%)

Balanced accuracy, 6 configurations. Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level.

Method	Condition	Metric	Value	People
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	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	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	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 ppC1-fz higher; 2 pp margin not met; gate passed.	51
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	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	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	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 ppC1-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 →

BOAS · Bitbrain Open Access Sleep dataset · One model, several questions · sleep (route 2)

Source page: One model, several questions · shared-representation-update.json

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%)

Balanced accuracy, 9 configurations. Dot: the estimate; line: 95% interval; dashed line, where drawn: chance level.

Method	Condition	Metric	Value	People
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	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	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	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 ppNo difference shown; equivalent within ±2 pp; gate passed.	100
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	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	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	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 ppC1-fz higher; 2 pp margin not met; gate passed.	100
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	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	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	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 ppNo 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 →

BOAS · Bitbrain Open Access Sleep dataset · Questions in language · sleep (route 3)

Source page: Questions in language · questions-in-language-update.json

Method	Condition	Metric	Value	People
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 ppNo difference shown; equivalent within ±2 pp; passes the triviality canary.	100
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 ppNo difference shown; equivalent within ±2 pp; passes the triviality canary.	100
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 ppNo difference shown; equivalent within ±2 pp; passes the triviality canary.	100
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 ppTPL(train) higher; 2 pp margin not met; passes the triviality canary.	100
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 ppDESC(train) higher; 2 pp margin not met; passes the triviality canary.	100
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–+3.967Unseen question. R 33.80; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+2.357Unseen question. R 7.04; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+1.047Unseen question. R 2.42; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+2.466Unseen question. R 9.65; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+1.698Unseen question. R 4.57; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+3.956Unseen question. R 31.30; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+1.744Unseen question. R 3.67; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
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–+1.856Unseen question. R 4.12; read-off leaves less error; 1.25 ratio margin not met; headroom gate passed.	100
CBraMod	P5 · label template − shuffled templates (TPL − SHUF) · unseen questions, implicit names · three seeds	Paired difference in AUROC	−0.378−0.501–−0.253Unseen question. SHUF higher; difference only: no margin; headroom gate passed.	100
CBraMod	P5 · label template − shuffled templates (TPL − SHUF) · unseen questions, explicit names · three seeds	Paired difference in AUROC	+0.629+0.603–+0.654Unseen 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 →

WBCIC-SHU motor imagery dataset · Later sessions · two-class motor imagery, session 1 to session 3

Source page: Later sessions · Chance level 50.0% · later-sessions-update.json

Method	Condition	Metric	Value	People
CBraMod · frozen encoder, session-1 ridge readout	Session 1 → session 3 · no session-3 labels	Balanced accuracy	67.7%64.8%–70.5%	51
CBraMod · frozen encoder, session-1 ridge readout	Minus the relative spectral ridge, the same people and trials	Paired difference	+13.8 pp+11.3 pp–+16.4 ppA 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 · Later sessions · three-class motor imagery, session 1 to session 3

Source page: Later sessions · Chance level 33.3% · later-sessions-update.json

Method	Condition	Metric	Value	People
CBraMod · frozen encoder, session-1 ridge readout	Session 1 → session 3 · no session-3 labels	Balanced accuracy	51.6%45.2%–58.3%	11
CBraMod · frozen encoder, session-1 ridge readout	Minus the relative spectral ridge, the same people and trials	Paired difference	+14.0 pp+7.5 pp–+20.8 ppA 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 → · All 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.)	Motor imagery & rest	not in the authors' published pretraining list (checked 2026-10-04)	source ↗
EEGMAT (PhysioNet EEG During Mental Arithmetic Tasks; Zyma et al. 2019)	Arithmetic & rest	not in the authors' published pretraining list (checked 2026-10-04)	source ↗
ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)	P300 target ERP	not in the authors' published pretraining list (checked 2026-10-04)	source ↗
ds005383 TMNRED (Bai et al. 2025, Sci. Data 12:701)	Semantic target ERP	not in the authors' published pretraining list (checked 2026-10-04)	source ↗
EESM19 Ear-EEG Sleep Monitoring 2019 (OpenNeuro ds005185; Mikkelsen et al.)	Sleep staging	not in the authors' published pretraining list (checked 2026-10-04)	source ↗
BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)	SSVEP · 8 channels · SSVEP · 4 channels	not in the authors' published pretraining list (checked 2026-10-04)	source ↗
ds005342 EEG offline/online MI for standing and sitting (Triana-Guzman et al.)	Idle & command	not in the authors' published pretraining list (checked 2026-10-04)	source ↗

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 (2026-10-08), later-sessions-update-20261008 (2026-10-08), shared-representation-update-20261007 (2026-10-07), reliable-decisions-update-20261004 (2026-10-04), deployment-topics-20260920-v1 (2026-09-20), 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. BibTeX for the site and its releases → · CITATION.cff ↗