TL;DR
The authors introduce eegdash-corpus-v0, a massive, standardized, and training-ready collection of 77 public EEG datasets totaling over 14,000 hours of recordings. By providing a unified, Zarr-based data store with pre-processed signals and clear licensing, they significantly lower the barrier to entry for researchers looking to build and benchmark the next generation of EEG foundation models.
What Problem Does This Paper Address?
Training "foundation models"—large-scale neural networks designed to be fine-tuned for various downstream tasks—requires an enormous amount of data. In the world of Electroencephalography (EEG), this is a major hurdle.
Currently, the largest EEG datasets are clinical archives that cannot be redistributed due to privacy and legal constraints. Publicly available data, while plentiful, is a "data engineering nightmare": it is scattered across different archives (like OpenNeuro and NEMAR), exists in a dozen different file formats, and uses inconsistent channel naming conventions and sampling rates. Assembling a clean, unified corpus for training can take months of manual effort. This paper aims to solve this "cold start" problem by releasing a curated, legally vetted, and pre-processed version of the public EEG commons.
Methods
The authors built the corpus by following a rigorous pipeline:
- Aggregation: Using the EEGDash platform, they fetched 77 redistributable datasets from OpenNeuro and NEMAR.
- Standardization (The "REVE Recipe"): To ensure the data is immediately useful, they applied the preprocessing steps used by REVE, a state-of-the-art EEG foundation model. This includes band-pass filtering (0.5–99.5 Hz), resampling to 200 Hz, and resolving electrode names to standard 3D positions.
- Storage: The data is stored in a Zarr store—a chunked, compressed, N-dimensional array format optimized for parallel cloud-native computing. This allows a training script to read only the specific samples it needs without loading massive files into memory.
- Validation: To prove the corpus works, they retrained the REVE-Base model from scratch using only this new public corpus and compared its performance against the original model (which was trained on a much larger, private clinical dataset).
Datasets
The eegdash-corpus-v0 consists of:
- 77 public datasets (56 from OpenNeuro, 21 from NEMAR).
- 14,075 hours of recordings from 7,272 subjects.
- 6.21 billion tokens per epoch (where a token is a 1-second patch of data from a single channel).
The corpus is dominated by the PEERS dataset, which accounts for 65% of the total hours. Other significant contributors include the Healthy Brain Network (HBN) and the NMT archive. For evaluation, the authors used OpenEEGBench, testing on seven public tasks ranging from motor imagery to seizure detection.
Key Insights and Ideas
- The "Token Density" Advantage: While the public corpus only contains about 23% of the hours used to train the original REVE model (which relied heavily on 19-channel clinical recordings), it contains roughly 79% of the tokens. This is because research datasets often use high-density caps (128+ channels), providing more "training signal" per hour of recording.
- Public Data is Sufficient for Feature Learning: Under "linear probing" (where the model's weights are frozen and only a small classification head is trained), the model trained on this public corpus performed almost identically to the one trained on 60,000+ hours of clinical data. This suggests that public data is diverse enough to learn robust, general-purpose EEG features.
- Licensing Transparency: Every "bucket" in the Zarr store includes the original dataset's license (e.g., CC0, CC-BY). This allows researchers to easily filter the corpus for specific legal requirements, such as excluding datasets that forbid commercial use.
Key Results
The authors successfully demonstrated that their "training-ready" corpus produces competitive models:
- Linear Probing: The retrained REVE-Base achieved a mean score of 0.331 across seven tasks, nearly matching the 0.333 of the original model trained on 4.4x more data.
- Fine-tuning: The retrained model reached 0.273, lagging slightly behind the original's 0.325. The authors attribute this gap partly to the original model's exposure to specific clinical patterns (like seizures) that are rarer in the public corpus.
- Efficiency: The patched Zarr loader is significantly faster than standard methods, enabling a training step on four GPUs in just 0.92 seconds.
Limitations and Future Work
- Fixed Recipe: The current version fixes the sampling rate (200 Hz) and window size (10s). Models requiring higher frequency information or different normalization might need to re-process the raw data.
- Paradigm Imbalance: The corpus is heavily skewed toward memory and cognition tasks (69% of hours). It lacks significant representation for sleep studies or intracranial recordings.
- Future Versions: The authors plan to expand the corpus with more non-gated datasets and provide alternative "recipes" (e.g., different frequency bands) to support a wider variety of model architectures like LaBraM or BIOT.
The corpus is available as a gated dataset on Hugging Face under EEGDash/eegdash-corpus-v0.
Summary updated