braindecode/STEEGFormer-largeV2

STEEGFormer (largeV2)

ViT-MAE EEG foundation model — braindecode port of ST-EEGFormer (largeV2 variant).

Provenance

• Weights ported from: LiuyinYang1101/STEEGFormer, release ST-EEGFormer-largeV2 (asset large_weights_only_210.pth).
• Upstream license: MIT. The braindecode wrapper code is BSD-3-Clause.
• The pre-trained encoder is loaded faithfully (numerical equivalence verified: pre-encoder bit-exact, post-encoder relative error ~4e-6). The MAE decoder is dropped and the classification head is re-initialised.

Architecture

	embed_dim	depth	num_heads	patch_size	channel vocab
largeV2	1024	24	16	16	256

This variant uses a 256-slot channel vocabulary (further pre-trained on HBN for the EEG 2025 Foundation Challenge). Name-based mapping via chs_info works for standard electrodes; pass chan_pos_idx explicitly for the HBN montage / non-standard channels.

Training data

The data that shaped these weights. The ids are the datasets: field of this card, which braindecode reads into model.card_data.datasets when you call from_pretrained (MOABB class names for MOABB datasets, OpenNeuro / NEMAR accessions, Hub ids otherwise). nm/on ids are datasets on NEMAR (on = NEMAR's copy of the OpenNeuro dataset with the same number). Rows marked private: are data the authors did not release; they are not in datasets:.

datasets: id	Dataset	Stage	Modality
nm000245	EEG-MI-BCI: Cho et al. 2017, EEG datasets for motor imagery BCI (52 subjects, 64 ch, left/right-hand MI; GigaDB 100295); MOABB Cho2017	pretraining	EEG
nm000172	High Gamma Dataset (HGD), Schirrmeister et al. 2017 (20 subjects, 128 ch, executed movements); MOABB Schirrmeister2017	pretraining	EEG
nm000139	BCI Competition IV dataset 2a (9 subjects, 22 ch, 4-class MI); MOABB BNCI2014_001	pretraining	EEG
nm000135	BCI Competition IV dataset 2b (9 subjects, 3 ch, 2-class MI); MOABB BNCI2014_004	pretraining	EEG
Kaya2018	Kaya et al. 2018, A large electroencephalographic motor imagery dataset for EEG BCIs (Sci. Data 5:180211; Figshare collection 3917698). Paper uses two subsets, 'Large-MI-Classic' and 'Large-MI-5F'. MOABB Kaya2018 (moabb>=1.7.2) loads only the 17 CLA files, not the 5F (or HaLT) files (NEMAR nm000137 holds a subset of it)	pretraining	EEG
Won2022	Won et al. 2022, EEG dataset for RSVP and P300 speller BCIs (Sci. Data 9:388; 55 participants P300 speller, 50 RSVP; Figshare collection 5769449); not in MOABB	pretraining	EEG
nm000129	BETA: large benchmark database toward SSVEP-BCI application, Liu et al. 2020 (70 subjects, 40 targets); MOABB Liu2020BETA	pretraining	EEG
nm000339	Stieger et al. 2021, continuous sensorimotor-rhythm BCI learning in a large population (62 subjects, ~600 h, 62 ch; Figshare 13123148); MOABB Stieger2021	pretraining	EEG
SparrKULee	SparrKULee: speech-evoked auditory response repository of KU Leuven (Bollens et al. 2023; 85 participants, 64 ch BioSemi); called 'KUL Auditory Decoding Dataset' in the paper (NEMAR nm000238 holds a subset of it)	pretraining	EEG
SEED-V	SEED-V emotion EEG dataset (Liu et al. 2022; 16 subjects x 3 sessions, 62 ch, five emotions)	pretraining	EEG
private:KU-Leuven-in-house-EEG	In-house EEG recorded in the authors' lab (KU Leuven Laboratory for Neuro- & Psychophysiology), multiple paradigms (e.g. MI and SSVEP) and diverse montages; not released, not itemised	pretraining	EEG
HBN	Healthy Brain Network EEG (HBN-EEG), as distributed for the NeurIPS 2025 EEG Foundation Challenge. Public releases R1-R11 on OpenNeuro: R1 ds005505, R2 ds005506, R3 ds005507, R4 ds005508, R5 ds005509, R6 ds005510, R7 ds005511, R8 ds005512, R9 ds005514, R10 ds005515, R11 ds005516 (eegdash const.py); R12 = withheld challenge test set. WHICH of these the authors used for the HBN pre-training is NOT documented.	pretraining	EEG

• Inherited ST-EEGFormer-large MAE pre-training (196 epochs per the ST-EEGFormer-large release). MAE pre-training corpus of App. E.2 (pp. 25-26), '11 public datasets' + an in-house set, as the paper lists them: 1) EEG-MI-BCI (Cho et al., 2017) -> Cho2017; 2) HGD (Schirrmeister et al., 2017) -> Schirrmeister2017; 3) BCI-Comp-IV2a (Tangermann et al., 2012) -> BNCI2014_001 (official training session only; the official test session was the validation split, E.4 p.26); 4) BCI-Comp-IV2b (Tangermann et al., 2012) -> BNCI2014_004; 5) Large-MI-Classic and 6) Large-MI-5F (both Kaya et al., 2018) -> Kaya2018 (one dataset, two paradigm subsets); 7) P300 (Won et al., 2022: P300 speller + RSVP) -> Won2022; 8) SSVEP (Liu et al., 2020 = BETA) -> Liu2020BETA; 9) Online MI BCI Classification (Stieger et al., 2021) -> Stieger2021; 10) KUL Auditory Decoding Dataset (Bollens et al., 2023 = SparrKULee) -> SparrKULee; 11) SEED-V (Liu et al., 2022) -> SEED-V; plus in-house EEG recorded in the authors' lab (multiple paradigms, e.g. MI and SSVEP) -> private:KU-Leuven-in-house-EEG. Preprocessing: notch, 0.1-64 Hz band-pass, 128 Hz, per-channel z-score (E.3). Windows: 6 s / 0.5 s hop for continuous sets, 6 s / 2.5 s hop for the three large MI sets, 2 s / 0.125 s hop for SSVEP (E.4). Validation: BCI-Comp-IV2a official test set; 20% held out for every other set; >8M overlapping segments overall.
• Further MAE pre-training on HBN-EEG for the NeurIPS 2025 EEG Foundation Challenge, channel vocabulary extended 145 -> 256 slots. Which HBN releases/tasks/subjects, sampling (the challenge data are the 100 Hz 'L100' downsampled releases; the authors' challenge model upsamples 100 -> 128 Hz) and how many epochs are not stated in any source we found. The challenge LoRA fine-tuning (eeg_foundation_2025/, eeg2025/KU_Leuven weights) is downstream of this checkpoint and NOT part of it.
• HBN: the source does not say which Healthy Brain Network EEG releases were used. The public releases are on NEMAR as on005505 (R1) to on005512 (R8) and on005514 (R9) to on005516 (R11), plus nm000103 (non-commercial release).
• Starts from the weights of braindecode/STEEGFormer-large, whose card lists the data those saw.
• Incomplete list. HBN releases used for the further pre-training are not documented (checked: STEEGFormer README + full git history, eeg_foundation_2025/ code and README, GitHub release notes, eeg2025/hall-of-fame-2025 KU_Leuven submission, HF eeg2025/KU_Leuven card, VSC blog post, challenge paper arXiv:2506.19141v3, eeg2025.github.io data/rules/faq). The challenge provided R1-R11 publicly (R5 = validation, R12 withheld test), so the HBN part is a subset of ds005505-ds005516 (minus ds005513) at most, but the exact subset is unknown. A dataset missing here is therefore not proof that these weights never saw it.
• Sources: Yang et al., ICLR 2026, https://openreview.net/forum?id=5Xwm8e6vbh; PDF via Wayback snapshot http://web.archive.org/web/20260609032700/https://openreview.net/pdf?id=5Xwm8e6vbh App. E.2, p.25; https://github.com/LiuyinYang1101/STEEGFormer README.md @542ee17 (section 7); https://github.com/LiuyinYang1101/STEEGFormer/releases/tag/ST-EEGFormer-largeV2 (release name); arXiv:2506.19141v3 (EEG Foundation Challenge), p.4.

Usage

from braindecode.models import STEEGFormer

model = STEEGFormer.from_pretrained(
    "braindecode/STEEGFormer-largeV2",
    n_outputs=4, n_chans=22, n_times=1000, chs_info=chs_info,
)
# Encoder features:  out = model(x, return_features=True); out["features"]

Citation

Yang, L., Sun, Q., Li, A. & Van Hulle, M. M. (2026). Are EEG foundation models worth it? Comparative evaluation with traditional decoders in diverse BCI tasks. ICLR 2026. https://openreview.net/forum?id=5Xwm8e6vbh