braindecode/meggpt-pretrained

MEG-GPT (pretrained, Cam-CAN)

Weights of MEG-GPT (Huang et al., Imaging Neuroscience 2026) and of the encoder of its tokenizer, EphysTokenizer, for braindecode.models.MEGGPT, converted from the authors' release to braindecode's parameter names. MEG-GPT is a causal time-attention GPT trained on tokenised source-space MEG (52 parcels at 250 Hz) from the Cam-CAN resting-state data. The class is added by braindecode#1555 (braindecode 2.1).

from braindecode.models import MEGGPT

model = MEGGPT.from_pretrained("braindecode/meggpt-pretrained", n_outputs=4)
features = model(x, return_features=True)["features"]

The classification head (final_layer, n_outputs=2 in config.json) is not part of the release: it is PyTorch's default init under torch.manual_seed(0). Passing n_outputs to from_pretrained replaces it. Fine-tune or linear-probe before use.

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
CamCAN	Cambridge Centre for Ageing and Neuroscience (Cam-CAN) MEG data, Shafto et al. 2014 / Taylor et al. 2017	pretraining, pretraining	MEG

• EphysTokenizer: paper trains it on the first 50 of 612 Cam-CAN subjects (eyes-closed rest, 52-parcel source space)
• MEG-GPT next-token pretraining on tokenised Cam-CAN resting-state; released config: 621 Cam-CAN train sessions, 90/10 per-subject window split
• Sources: arXiv:2510.18080v1 Sec. 2.5 Datasets; MEG-GPT@78ff9439 models/meg-gpt/config.yaml; MEG-GPT@78ff9439 examples/camcan/01_train_tokenizer.py.

Intended input

x is (batch, 52, n_times) float, source-space MEG: the 52 parcels of the released parcellation (Glasser 52) in its order, sign-flip aligned and z-scored per parcel over the recording (tokenize_session(standardize=True) in the reference; Section 2.5.3 of the paper), at 250 Hz. The model does not normalise its input and does not check sfreq. The window is cut into 100-sample segments; set on_non_divisible for windows that are not a multiple of 100.

Source

• Paper: R. Huang, S. Cho, C. Gohil, O. Parker Jones, M. Woolrich. MEG-GPT: A transformer-based foundation model for magnetoencephalography data. Imaging Neuroscience 4, IMAG.a.1301 (2026). https://doi.org/10.1162/imag.a.1301 (arXiv:2510.18080)
• Code and weights: OHBA-analysis/MEG-GPT @ 78ff9439febaeb8e838fd670ec589578e635fd0c (MIT), directory models/; tokenizer code: OHBA-analysis/EphysTokenizer @ 44cac87caee7ae2521d2b88e7bc71c510b10a023 (MIT).

original file (MEG-GPT @ 78ff9439)	sha256
models/meg-gpt/model_state.pt	10755496172a8c921cd9505884d00c6370bed211e99863471e790f1bbea2eeb9
models/meg-gpt/config.yaml	4434140a4aff5f19124a257cf0425fb9b564b05060c85317f31a8f935ac91069
models/tokenizer/model_state.pt	1d766a574942122ed28d2d08dcd2e33dc53b2b8a30628c100d786516152cf2a7
models/tokenizer/vocab.pkl	ea0273f3efc6f349ae231b00332a511f303e13aadd69e18c3c6b93ea65838959
models/tokenizer/config.yaml	f312ab9a56d3a0f2937afd6a9235e6af87887f808be054103634bcd3367b6241

Conversion

convert.py (in this repository) was run with braindecode at bruAristimunha/braindecode@2f253bc1d71b9cba66ad40e90907b7ac12cec343 (branch feat/meg-meggpt, PR #1555):

python convert.py <MEG-GPT>/models <out_dir>

It renames the keys of the two state dicts, stores the vocabulary (vocab.pkl["label_map"]) as the tokenizer.vocab buffer, loads them into MEGGPT(n_chans=52, n_outputs=2, sfreq=250) and calls save_pretrained. Tensor values are copied unchanged.

Dropped: the Cam-CAN session embedding (input_embedding_layer.extra_embeds.0.*, 621 sessions; specific to the pretraining data, and the paper discards it for downstream use, Section 2.5.4), the next-token prediction_head, the attention-mask buffers (rebuilt from the configuration, checked equal), the tokenizer's signal decoder (token_basis_layer) and its annealing temperature.

file	sha256
model.safetensors	e009d3f4d81eb6d194f04a72f59cca19140600266ab7d464691f77d388ccd821
pytorch_model.bin	3e5f7116d3d1719fc74925d2f387bb4b089c16c895d8d4af606c406570b175b1
config.json	f022492a654181df38c66f9230b681fa20f521039de83d6c119dfead4f61bc5f
convert.py	792b681d01dd3e3e2a16dbd51f67cfc80da3387094714ed9f9b39cb28e657172

Verification

MEGGPT.from_pretrained(<this repo>, strict=True) against the original code (CPU, float32, eval mode, same inputs): token ids identical; decoder hidden states, time-averaged features, logits through the released prediction head and the gradients of 99 tensors in one backward pass agree with max-abs difference 0.0.

License

MIT, as the original release: MEG-GPT licence (Copyright (c) 2026 OHBA Analysis Group). The weights ship in models/ of that MIT-licensed repository, with no separate weight licence. Converted and re-hosted with attribution to the authors; for corrections or removal, open an issue at https://github.com/braindecode/braindecode/issues.

Citation

@article{huang2026meggpt,
  title   = {{MEG-GPT}: A transformer-based foundation model for magnetoencephalography data},
  author  = {Huang, Rukuang and Cho, SungJun and Gohil, Chetan and Parker Jones, Oiwi and Woolrich, Mark},
  journal = {Imaging Neuroscience},
  volume  = {4},
  pages   = {IMAG.a.1301},
  year    = {2026},
  doi     = {10.1162/imag.a.1301}
}

@article{aristimunha2025braindecode,
  title   = {Braindecode: a deep learning library for raw electrophysiological data},
  author  = {Aristimunha, Bruno and others},
  journal = {Zenodo},
  year    = {2025},
  doi     = {10.5281/zenodo.17699192},
}