🧠 Open EEG Bench

Benchmark any EEG foundation model with one function call.

✨ Why Open EEG Bench?

• 🎯 One function, all results — Call benchmark() with your model and get a full evaluation across 12 datasets
• 📦 Zero preprocessing — All datasets are pre-windowed and hosted on HuggingFace Hub, ready to use
• ⚡ 8 fine-tuning strategies — Frozen linear probing, ridge probing, LoRA, IA3, AdaLoRA, DoRA, OFT, and full fine-tuning
• 🔌 Bring your own model — Any PyTorch model that takes EEG input and returns features works out of the box
• 🔒 Reproducible by design — A single config object fully describes a run. No YAML files, no hidden state

🚀 Installation

pip install open-eeg-bench

🏁 Benchmark your model

import open_eeg_bench as oeb

results = oeb.benchmark(
    model_cls="my_package.MyModel",  # import path to your model class
    checkpoint_url="https://my-weights.pth",
)

print(results)  # pd.DataFrame

This runs linear probing on all 12 datasets and returns a DataFrame with one row per result.

You can pick specific datasets, fine-tuning strategies, classification heads and number of initialization seeds:

results = oeb.benchmark(
    model_cls="my_package.MyModel",
    checkpoint_url="https://my-weights.pth",
    datasets=["arithmetic_zyma2019", "bcic2a", "physionet"],
    finetuning_strategies=["frozen", "lora"],
    peft_target_modules=[  # necessary for LoRA, IA3, AdaLoRA, OFT, and DoRA
        "encoder.linear1", 
        "encoder.linear2"
    ],
    heads=["linear_head", "mlp_head"],
    n_seeds=5,
)

Need to run on a SLURM cluster? No sbatch scripts needed — see Running on a cluster.

📐 Model requirements

Your model only needs to:

1. Accept input of shape (batch, n_chans, n_times)
2. Return output of shape (batch, n_outputs)
3. Have a named module for the classification head (default: self.final_layer)

⚙️ benchmark() parameters

📊 Available datasets

All 12 datasets are pre-windowed and hosted on HuggingFace Hub:

Preprocessing: Window lengths are dataset-dependant (see table above). All datasets are high-pass filtered at 0.1 Hz, except for tasks with short trial windows (2 s or less), where we use 0.5 Hz. All datasets are resampled to 100 Hz. Model-specific normalization (e.g. z-scoring) can be applied via the normalization parameter.

🧩 Available fine-tuning strategies

📄 License

BSD-3-Clause (see LICENSE.txt)

📣 Citation

If you use Open EEG Bench in your research, please cite it. Machine-readable metadata is provided in CITATION.cff — used by GitHub's "Cite this repository" button, reference managers, and Zenodo when archiving releases.

@software{open_eeg_bench,
  title     = {Open EEG Bench: Benchmarking parameter-efficient fine-tuning of EEG foundation models},
  author    = {Guetschel, Pierre and Aristimunha, Bruno and Truong, Dung and Kokate, Kuntal and Moreau, Thomas and Tangermann, Michael and Delorme, Arnaud},
  year      = {2026},
  url       = {https://github.com/braindecode/OpenEEGBench},
  doi       = {10.5281/zenodo.19698863}
}

You may also be interested in the following articles which use Open EEG Bench:

@inproceedings{guetschel2026openeegbench,
  title        = {Toward {OpenEEG-Bench}: A Live Community-Driven Benchmark for {EEG} Foundation Models},
  author       = {Guetschel, Pierre and Aristimunha, Bruno and Truong, Dung and Kokate, Kuntal and Tangermann, Michael and Delorme, Arnaud},
  booktitle    = {Proceedings of the 34th European Signal Processing Conference (EUSIPCO 2026)},
  year         = {2026},
  address      = {Bruges, Belgium},
  month        = aug,
  pages        = {1--5},
  organization = {EURASIP},
}

📚 Further reading

• Running on a cluster — SLURM submission, infra options, caching, execution modes
• Running with Benchopt — Alternative CLI-based workflow using the Benchopt framework
• Advanced usage — Experiment class, architecture overview, braindecode integration
• Built-in backbones — Pretrained models shipped with the benchmark
• Contributing — How to add backbones, datasets, or fine-tuning strategies
• References — Citations and key libraries