ST-EEGFormer: results on public EEG datasets

Foundation model

ST-EEGFormer: results on public EEG datasets

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; EEGMAT; BETA; ds006593; TMNRED / ds005383; EESM19 scalp subset; ds005342 · 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 ST-EEGFormer Base · ST-EEGFormer Large · STEEGFormer

Reference ST-EEGFormer, OpenReview 5Xwm8e6vbh ↗

Directory status: Evaluated · Official code ↗ · Description sources openreview.net · github.com

Published results

Grouped by dataset. Compare figures within a group only: across groups the task, cohort, chance level and electrode layout all change.

ds003810 · v9 foundation encoders, frozen · motor imagery & rest

Read with its protocol: Core-matrix protocol: Motor imagery & rest · Chance level 50.0% · foundation-models-mi-rest.csv

ST-EEGFormer Base

58.1% (53.7%–63.3%)

ST-EEGFormer Large

62.4% (57.4%–67.8%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	58.1%53.7%–63.3%	10
ST-EEGFormer Large	Frozen encoder + ridge head	Balanced accuracy	62.4%57.4%–67.8%	10

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

EEGMAT · v9 foundation encoders, frozen · arithmetic & rest

Read with its protocol: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · foundation-models-arithmetic-rest.csv

ST-EEGFormer Base

64.5% (59.0%–70.0%)

ST-EEGFormer Large

63.4% (57.8%–69.0%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	64.5%59.0%–70.0%	36
ST-EEGFormer Large	Frozen encoder + ridge head	Balanced accuracy	63.4%57.8%–69.0%	36

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

EEGMAT · Model adaptation, v9 · new people, same task, one fixed recipe (not a ranking)

Read with its protocol: Core-matrix protocol: Arithmetic & rest · Chance level 50.0% · foundation-models-update.json

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%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + trained head	Balanced accuracy	63.2%59.7%–66.7%1,538 trainable parameters.	36
ST-EEGFormer Base	LoRA rank 4 + head	Balanced accuracy	65.1%61.0%–69.4%148,994 trainable parameters.	36
ST-EEGFormer Base	LoRA minus frozen + head, same people and folds	Paired difference	+1.9 pp−0.2 pp–+4.1 pp20 people improved, 16 got worse, 0 unchanged. The interval includes zero: no change is established.	36

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

BETA · v9 foundation encoders, frozen · SSVEP, 8 channels

Read with its protocol: Core-matrix protocol: SSVEP · 8 channels · Chance level 2.5% · foundation-models-beta-8ch.csv

ST-EEGFormer Base

17.1% (15.5%–18.8%)

ST-EEGFormer Large

22.6% (20.1%–25.3%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	17.1%15.5%–18.8%Pretraining exposure: in the authors' published pretraining list.	70
ST-EEGFormer Large	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. No model’s authors endorse these results. Every weights licence and its terms →

BETA · v9 foundation encoders, frozen · SSVEP, 4 channels

Read with its protocol: Core-matrix protocol: SSVEP · 4 channels · Chance level 2.5% · foundation-models-beta-4ch.csv

ST-EEGFormer Base

17.5% (15.8%–19.3%)

ST-EEGFormer Large

22.4% (19.9%–25.0%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	17.5%15.8%–19.3%Pretraining exposure: in the authors' published pretraining list.	70
ST-EEGFormer Large	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. No model’s authors endorse these results. Every weights licence and its terms →

ds006593 · v9 foundation encoders, frozen · P300 target ERP

Read with its protocol: Core-matrix protocol: P300 target ERP · Chance level 50.0% · foundation-models-p300-target.csv

ST-EEGFormer Base

53.6% (52.3%–54.8%)

ST-EEGFormer Large

52.7% (51.3%–54.2%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	53.6%52.3%–54.8%	21
ST-EEGFormer Large	Frozen encoder + ridge head	Balanced accuracy	52.7%51.3%–54.2%	21

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

TMNRED / ds005383 · v9 foundation encoders, frozen · semantic target ERP

Read with its protocol: Core-matrix protocol: Semantic target ERP · Chance level 50.0% · foundation-models-semantic-target.csv

ST-EEGFormer Base

54.6% (52.9%–56.4%)

ST-EEGFormer Large

57.8% (56.0%–59.6%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	54.6%52.9%–56.4%	30
ST-EEGFormer Large	Frozen encoder + ridge head	Balanced accuracy	57.8%56.0%–59.6%	30

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

EESM19 scalp subset · v9 foundation encoders, frozen · sleep staging

Read with its protocol: Core-matrix protocol: Sleep staging · Chance level 20.0% · foundation-models-sleep-scalp.csv

ST-EEGFormer Base

77.3% (75.6%–78.8%)

ST-EEGFormer Large

79.6% (78.2%–81.0%)

Balanced accuracy, 2 configurations. Dot: the estimate; line: 95% interval; dashed line: chance level where the payload records one.

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + ridge head	Balanced accuracy	77.3%75.6%–78.8%	20
ST-EEGFormer Large	Frozen encoder + ridge head	Balanced accuracy	79.6%78.2%–81.0%	20

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

ds005342 · v9 foundation encoders, frozen · idle & command

Read with its protocol: Core-matrix protocol: Idle & command · foundation-models-idle.csv

Method	Condition	Metric	Value	People
ST-EEGFormer Base	Frozen encoder + linear head	Commands detected ≤3 s	16 / 601 of 4 people always abstained.	4
ST-EEGFormer Base	Frozen encoder + linear head	Idle false activations	1 / 60	4
ST-EEGFormer Large	Frozen encoder + linear head	Commands detected ≤3 s	20 / 601 of 4 people always abstained.	4
ST-EEGFormer Large	Frozen encoder + linear head	Idle false activations	1 / 60	4

Weights terms: ST-EEGFormer under MIT. No model’s authors endorse these results. Every weights licence and its terms →

All methods → · All datasets →

Checkpoints, inputs and terms

What was run, as the v9 release records it: each checkpoint’s revision and SHA-256, its encoder parameters, the footnote its rows carry and its weights licence. No weights, adapted weights or LoRA deltas were kept or shared, and no endorsement by the model’s authors is implied.

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 ↗
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 ↗
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 and linked to the source. A dataset absent from the list is not proof that its recordings were never seen; recording-level audits were not done.

ST-EEGFormer Base · ST-EEGFormer Large

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 ↗ · 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 ↗ · source ↗
ds006593 cBCI Matrix Multimodal Dataset (Celik et al.)	P300 target ERP	not in the authors' published pretraining list (checked 2026-10-04)	source ↗ · 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 ↗ · 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 ↗ · source ↗
BETA 40-target SSVEP (Liu et al. 2020, Front. Neurosci. 14:627)	SSVEP · 8 channels · SSVEP · 4 channels	in the authors' published pretraining list	source ↗ · 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 ↗ · source ↗

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 (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. BibTeX for the site and its releases → · CITATION.cff ↗