BCI ReportFoundation 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; EEGMAT; BETA; ds006593; TMNRED / ds005383; EESM19 scalp subset; ds005342; Mobile BCI dataset (SSVEP and ERP paradigms) · 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 LaBraM · Large Brain Model · LaBraM-Base · Labram · Jiang et al. 2024
Directory status: Evaluated · Official code ↗ · Description sources arxiv.org · github.com · github.com
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.
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | Rest versus right-hand imagery | Balanced accuracy | 53.5%51.4%–55.4% | 10 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | Serial subtraction versus resting EEG | Balanced accuracy | 64.6%60.5%–68.6% | 36 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | Head only · encoder frozen | Balanced accuracy | 56.6%54.2%–59.0% | 36 |
| LaBraM | Last block + head | Balanced accuracy | 65.7%61.6%–69.7% | 36 |
| LaBraM | LoRA rank 4 + head | Balanced accuracy | 64.1%60.2%–67.8% | 36 |
| LaBraM | LoRA rank 4 + head minus head only | Paired difference | +7.5 pp+4.8 pp–+10.1 pp | 36 |
| LaBraM | Last block + head minus head only | Paired difference | +9.1 pp+6.0 pp–+12.2 pp | 36 |
| LaBraM | LoRA rank 4 + head minus last block + head | Paired difference | −1.6 pp−4.4 pp–+1.0 pp | 36 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM · head only, encoder frozen | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 1 / 2,16035 of 36 people had nothing accepted. | 36 |
| LaBraM · head only, encoder frozen | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +0.6 pp−0.5 pp–+2.7 ppNo difference resolved: the interval includes zero. | 36 |
| LaBraM · LoRA rank 4 + head | Fixed threshold: calibrated confidence at least 0.8 | Windows accepted | 44 / 2,16024 of 36 people had nothing accepted. | 36 |
| LaBraM · LoRA rank 4 + head | Learned reject option minus calibrated confidence, same classifier | Error among the 80% most certain, difference | +1.9 pp+0.6 pp–+3.5 ppDifference 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). 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 →
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | 40 visual targets · 8 posterior electrodes | Balanced accuracy | 10.8%9.5%–12.2% | 70 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | 40 visual targets · 4 posterior electrodes | Balanced accuracy | 12.9%11.2%–14.6% | 70 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | Visual target versus nontarget events | Balanced accuracy | 49.4%47.7%–51.2% | 21 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | Reading · target versus nontarget events | Balanced accuracy | 53.2%51.3%–54.9% | 30 |
| Method | Condition | Metric | Value | People |
|---|---|---|---|---|
| LaBraM | Five stages · balanced scalp-EEG sample | Balanced accuracy | 71.1%69.0%–73.2% | 20 |
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.
Note on the check: SingLEM and LaBraM lists are matched by dataset name and reference; most rows carry no accession number.
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.
Pan-Junyu/LaBraM-at-Scale @ a0b0ce27 · quark/quark-base.safetensors3ffd4ca8fc7acf63dd89a8fc4be8c5d79666818e62bc61b10693767f0925648ePan-Junyu/LaBraM-at-Scale @ a0b0ce27 · quark/quark-large.safetensors8729eea4bb16567b11ce1d86d9e23b485324f4b5e8ad1122964003f7c89cd339Pan-Junyu/LaBraM-at-Scale @ a0b0ce27 · quark/quark-huge.safetensors246b875445b156d54e6ebb0b97f90dbd6890e848e8255290e7d9c7c48a2728c5Pan-Junyu/LaBraM-at-Scale @ a0b0ce27 · quark/quark-giant.safetensorsfe75959108ba07f0b435898dc43ceefca1de6b92b4f77c0486a6bec670ed835ePan-Junyu/LaBraM-at-Scale @ a0b0ce27 · atom/atom-large.safetensorsa28e97f7bd73bd7ff55f3b5d680f48807c335cf3134d1281228a061056aaf52bPan-Junyu/LaBraM-at-Scale @ a0b0ce27 · molecule/molecule-large.safetensorsaa590160283a5ba5a081ee89f1535f1ade584ffc27455d409ab277be3b4ac8d0Their cells, the Quark ladder and the Quark-G step are on “Does pretraining help EEG foundation models like LaBraM and CBraMod?”; 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 and, for the architecture, LaBraM (arXiv 2405.18765). No weights are distributed.
foundation-checkpoints-update.json · foundation-checkpoints-update-20261106
BCI Report (2026). LaBraM: results on public EEG datasets. https://bci.report/methods/labram/
Figures from releases foundation-checkpoints-update-20261106 (2026-11-06), reliable-decisions-update-20261004 (2026-10-04), adaptation-update-20261001 (2026-10-01), 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 ↗