Introduction#
What it is#
EEGDash Corpus is an aggregated, training-ready collection of public EEG datasets for pretraining foundation models. It is served in two forms: the raw datasets, each reachable through EEGDash by its identifier, and a converted Zarr store, built by applying one recipe once (200 Hz, a 0.5–99.5 Hz band, 10-second samples, microvolts left un-normalised, with each session’s mean and standard deviation and each channel’s position stored beside the samples), that a trainer reads directly. Zarr is the chunked array format the store is served in: a trainer reads the samples it needs, over the network or from a local copy, without unpacking the store.
Why#
The data behind EEG foundation models is fragmented across archives, file formats, sampling rates, montages and licences. Each group collates, checks, licenses and converts its own copy, and that work is not released in a form another model can train on. The corpus does it once: a group can start from the data rather than from the collation, with the licence of every dataset accounted for.
Starting point: REVE#
The first versions cover the corpus of REVE, the largest published EEG foundation model. REVE’s paper names 89 datasets; the release card of its public subset lists 70 entries, 4 of which the paper does not name, so 93 datasets are tracked in all. Version 0 is REVE’s public release rebuilt from the archives (66 datasets, 45,549 hours). Version 1 extends it to every publicly usable dataset named by REVE’s paper (78 datasets: 11 ingested by EEGDash and waiting for conversion, 1 still to be ingested); gated datasets are excluded, and datasets whose licences block a converted copy are listed with the reason. Later versions add public datasets beyond REVE’s, chosen by licence and usability.
What a version is#
A frozen table of datasets and the converted Zarr store built from it. Versions are additive; each carries its own table, provenance and conversion report.
How to get it today#
EEGDashDataset(dataset="<id>", cache_dir=...) for any raw dataset of the table;
the converted Zarr store is to be hosted through EEGDash.
How it was built and checked#
Fetch. OpenNeuro and NEMAR datasets through EEGDash into a cache that mirrors the archives.
Convert. The recipe above, per recording, with a report of what was dropped and why.
Verify. two datasets compared against the authors’ own loader sample by sample; REVE-Base retrained on the Zarr store scores within 0.011 (linear probe) and 0.024 (fine-tuning) of the released weights on eight public OpenEEGBench tasks.
Where the numbers come from#
Every cell carries a provenance label: zarr (measured in the converted Zarr store),
eegdash (EEGDash’s catalogue records), card (REVE’s release card), stated
(REVE’s paper), download (our copy). A bucket is one exact channel set of one
dataset in the Zarr store. Status vocabulary: Ingested and Converted (v0) · Ingested, but
not Converted (v1) · To Be Ingested (v1) · Blocked (License) · Gated (To Be Ingested)
· Excluded.
The corpus in numbers#
Licences#
Every dataset tracked, by licence, coloured by status: the licence decides whether a converted copy may be redistributed, and under which terms.