CORTEG, ST-EEGFormer small backbone

The starting point of CORTEG (Yang et al., 2026), ST-EEGFormer small adapted to intracranial recordings, for braindecode.models.CORTEG: the ST-EEGFormer small encoder (checkpoint-300.pth, sha256 98d2848abdaea7a143d399c07f67f8782e07dd66e7c309ddb058c7158a8a8b02, through its conversion braindecode/STEEGFormer-small at revision 633c9c0d), 25.3M parameters. The high-gamma patch embedding, which the scalp model does not have, is warm-started from the low-frequency one, linearly interpolated from 16 to 25 samples, as the authors' build_model does (LiuyinYang1101/CORTEG, experiments/run_regression_hilo_clean.py); it matches theirs to 0.0e+00 in float32 on CPU. convert_corteg_checkpoint.py in this repository reproduces the conversion.

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:.

• MAE self-supervised pre-training, 300 epochs (release note). 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.
• 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/releases/tag/ST-EEGFormer-small (release body).

Scope of this warm-start

This checkpoint inherits the scalp-EEG MAE corpus above from braindecode/STEEGFormer-small. No Stanford iEEG training has occurred here. The high-gamma patch embedding is warm-started from the low-frequency embedding; the spatial adapter and task readout are initialized on load and require downstream training.

Usage

from braindecode.models import CORTEG

model = CORTEG.from_pretrained(
    "braindecode/corteg-small-pretrained", n_outputs=5, n_times=128, chs_info=raw.info["chs"]
)

The spatial adapter, the head and, with merge_strategy="layerwise_gate", the gate are not in these files: Braindecode initializes them on load, the adapter from the electrode positions of chs_info (metres). The paper then trains the adapter, the high-gamma embedding, LoRA (r=4, alpha=16) on the last four blocks with their LayerNorms, and the head, with the rest frozen. See braindecode/corteg-stanford-pooled for the authors' trained Stanford model and the input streams.

Limitations

These files hold no trained intracranial component, so the model needs training before prediction. The check above covers the float32 CPU tensors of the warm start.

Citation

@inproceedings{yang2026corteg,
  title     = {{CORTEG}: Foundation Models Enable Cross-Modality Representation Transfer
               from Scalp to Intracranial Brain Recordings},
  author    = {Yang, Liuyin and Sun, Qiang and Van Dyck, Bob and Calvo Merino, Eva and
               Van Hulle, Marc M.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}

License

The ST-EEGFormer weights are released by Liuyin Yang under the MIT licence (LICENSE). Braindecode's code is BSD-3-Clause.

Cite Braindecode

If you use this braindecode port in scientific work, cite the software on Zenodo, doi:10.5281/zenodo.17699192, and the original model and datasets cited above.

@software{braindecode,
  author = {Aristimunha, Bruno and
            Guetschel, Pierre and
            Wimpff, Martin and
            Gemein, Lukas and
            Rommel, Cedric and
            Banville, Hubert and
            Sliwowski, Maciej and
            Wilson, Daniel and
            Brandt, Simon and
            Gnassounou, Théo and
            Paillard, Joseph and
            {Junqueira Lopes}, Bruna and
            Sedlar, Sara and
            Moreau, Thomas and
            Chevallier, Sylvain and
            Gramfort, Alexandre and
            Schirrmeister, Robin Tibor},
  title = {Braindecode: toolbox for decoding raw electrophysiological brain data
           with deep learning models},
  url = {https://github.com/braindecode/braindecode},
  doi = {10.5281/zenodo.17699192},
  publisher = {Zenodo},
  license = {BSD-3-Clause},
}