tankalapavankalyan/eeg-corpus-manifest

EEG Canonical Corpus Index

A queryable index over 79,313 hours of lossless EEG in a single unified .eegz (Zarr v3) format. The index ships as metadata only; the actual signal bytes live as canonical Zarr stores on S3 and are read on demand.

Recordings (canonical, ready to load)	185,802
Hours of EEG (canonical)	79,313.6
Canonical on-disk size	2.81 TB
Subjects	43,114
Source datasets indexed	411
Apples-to-apples compression vs source EEG (rows with canonical_size_bytes set)	3.05×

Every canonical row points at a Zarr-v3 group at s3://eeg-corpus-139156132535/canonical/v1/<dataset_id>/<recording_id>.eegz/ with FLAC-compressed int16 signal + parallel channel / events / annotations arrays. Open any one through manifest.eegz.open_recording or directly with zarr.open(canonical_uri) and get the same on-disk shape regardless of source format (.edf / .bdf / .set / .vhdr / .fif / .cnt / .gdf).

The original source files are kept at s3://eeg-datasets-139156132535/raw/ for provenance + the 0.31% residue that did not round-trip (see Non-OK recordings).

What's in here

The bundle is a star schema of four Parquet tables. The Dataset Viewer above lets you browse each one interactively; you can also load them programmatically.

Table	Rows	Description
datasets	411	One row per source dataset. License, DOI, paradigm summary, modality, paradigm categories.
recordings	186,386	The fact table. One row per recording. canonical_uri points at the .eegz store for the 185,802 ok rows; archival_uri is provenance to the original source file.
channels	11,530,668	One row per channel per recording. Name, type, units, electrode status, and 3D coordinates (x, y, z) when the dataset ships an electrodes.tsv.
subjects	43,114	One row per hashed subject. Age, sex, clinical status, handedness. The hash is per-dataset (cross-dataset subject dedup is a separate concern).

Round-trip contract

The canonical writer asserts one of two contracts per recording before upload:

roundtrip_class	Source formats	Contract	Ok rows
bit_exact	.edf, .bdf	Every digital sample equals the source byte-for-byte (within ≤1 LSB env-rounding tolerance).	62,157
near_lossless	.set, .vhdr, .fif, .cnt, .gdf	Per-channel reconstruction error ≤ 0.5 LSB of the digital→physical scale. Half-LSB is typically ~0.005 µV — well below the EEG noise floor.	35,908
(NULL)	(legacy)	Roundtrip class not recorded — the rows pre-date when the writer started persisting the class to the conversion ledger. The store itself passed the contract at write time.	87,737

The conversion ledger persisted per-store metadata including canonical_sha256 (SHA-256 of _eegz_integrity.json) on 2,402 stores; re-hashing those stores byte-for-byte returns 2,402 / 2,402 matches. 185,802 / 185,802 ok stores have a valid root zarr.json and signal/zarr.json on S3.

Loading one recording

from manifest.eegz import open_recording

uri = "s3://eeg-corpus-139156132535/canonical/v1/hbn_eeg/<recording_id>.eegz/"
rec = open_recording(uri)
print(rec.metadata.n_channels, rec.metadata.sampling_rate_hz, rec.metadata.duration_s)

# Lazy random-access slice — a single S3 GET fetches the relevant FLAC shard.
window = rec.signal[:, 60_000:120_000]   # samples 60k..120k across all channels

Or with pure Zarr:

import zarr
store = zarr.open("s3://eeg-corpus-139156132535/canonical/v1/<dataset>/<rid>.eegz/", mode="r")
signal = store["signal"][:]              # (n_channels, n_samples)
sfreq  = store.attrs["sampling_rate_hz"]
chnames = store["channels/name"][:]

Querying the manifest

Pandas

import pandas as pd
recordings = pd.read_parquet(
    "hf://datasets/tankalapavankalyan/eeg-corpus-manifest/recordings.parquet"
)
ok = recordings[recordings["conversion_status"] == "ok"]
hours_per_dataset = ok.groupby("dataset_id")["duration_s"].sum() / 3600
print(hours_per_dataset.sort_values(ascending=False).head())

DuckDB (joins across all four tables, zero downloads beyond columns you touch)

import duckdb
db = duckdb.connect()
db.sql("INSTALL httpfs; LOAD httpfs;")
BASE = "https://huggingface.co/datasets/tankalapavankalyan/eeg-corpus-manifest/resolve/main"

# Recordings ready to load, ordered by canonical store size.
db.sql(f"""
    SELECT dataset_id, recording_id, canonical_uri, roundtrip_class,
           canonical_size_bytes / 1e6 AS canonical_mb
    FROM '{BASE}/recordings.parquet'
    WHERE conversion_status = 'ok'
    ORDER BY canonical_size_bytes DESC NULLS LAST
    LIMIT 10
""").show()

Polars (lazy, S3 URIs preserved)

import polars as pl
rec = pl.scan_parquet(
    "hf://datasets/tankalapavankalyan/eeg-corpus-manifest/recordings.parquet"
)
(rec.filter((pl.col("conversion_status") == "ok") & (pl.col("sampling_rate_hz") >= 250))
    .group_by("dataset_id")
    .agg(pl.col("duration_s").sum().alias("seconds"),
         pl.col("recording_id").count().alias("n_rec"))
    .collect())

datasets

from datasets import load_dataset

ds = load_dataset("tankalapavankalyan/eeg-corpus-manifest", "recordings", split="train")
print(ds[0])
print(ds.features)

Schema — recordings (the main fact table)

All identifier columns are stable across builds via UUIDv5 on the archival URI.

Column	Type	Notes
recording_id	string	UUIDv5 over archival_uri; stable across builds
dataset_id	string	FK → datasets.dataset_id
canonical_uri	string	s3://.../canonical/v1/<dataset_id>/<recording_id>.eegz/ — Zarr v3 store. NULL when conversion_status != 'ok'.
canonical_size_bytes	int64	On-disk size of the .eegz store (sum of chunk + metadata bytes). NULL on legacy stores enriched from S3 listings rather than ledgers.
canonical_sha256	string	SHA-256 of the store's _eegz_integrity.json blob. NULL on stores written before the integrity-blob scheme shipped.
conversion_pipeline_id	string	Pipeline version reported by the converter.
conversion_status	string, never NULL	ok (185,802) / error_read (568) / error_download (9) / error_verify (3) / error_write (1) / not_yet_converted (3).
roundtrip_class	string	bit_exact / near_lossless / NULL (legacy).
conversion_dt_utc	timestamp[UTC]	When the conversion finished.
subject_id_in_dataset	string	As in source (sub-NDARxxx, sub-LTPxxx, anon clinical code).
subject_canonical_hash	string	FK → subjects.subject_hash.
session_id, run_id, task	string	BIDS-style entities; nullable.
archival_uri	string	s3://.../raw/... pointer to the source file. Provenance; reading from here only makes sense for non-ok residue.
archival_format	string	edf / bdf / set / vhdr / fif / cnt / gdf.
file_size_bytes	int64	Source file size from S3 listing; always populated.
duration_s	float64	Aligned to canonical zarr.json for ok rows; inherited from BIDS sidecar otherwise.
n_channels, n_eeg_channels	int32	Aligned to canonical zarr.json for ok rows.
sampling_rate_hz	float32	Aligned to canonical zarr.json for ok rows.
reference	string	e.g. Cz, common.
montage_name	string	e.g. HydroCel Geodesic Sensor Net.
channel_names	list[string]	Inline, per recording.
manufacturer	string	EGI, BioSemi, … (sparse).
power_line_frequency	float32	50 / 60 Hz.
recording_type	string	continuous / epoched.
modality	string	eeg (room for ieeg, meg in future).
bids_compliant	bool
bids_entities, companion_uris	JSON string	Stringified map<string, string>; parse with json.loads.
header_source, header_read_status	string	Provenance for the original BIDS-sidecar / EDF-header read.

For converted recordings, sampling_rate_hz / duration_s / n_channels / n_eeg_channels in recordings.parquet and the corresponding attributes on the canonical store's zarr.json are byte-for-byte identical across all 185,802 ok rows.

See the full schema in manifest/schema.py on the companion code repo.

Non-OK recordings

The 584 non-ok rows (0.31% of the corpus) break down as:

Count	Dataset(s)	Reason	Recoverable?
226	openneuro_ds002181	.fdt binary samples never uploaded by the publisher	❌ permanent — nothing on S3 to convert from
85	openneuro_ds004166	corrupt .set metadata; EEGLAB reader handled 128/213	⚠️ partial — needs reader extension
254	scattered across 42 small datasets (1–39 each)	upstream source quirks; mostly corrupt .set metadata, missing companions, MNE-unreadable epoched files	⚠️ case-by-case
3	openneuro_ds004952	read OK but failed signal-level verify	⚠️ inspect why
9	various	S3 transient error_download	✅ retry
3	peers_memory BDFs	shard worker terminated before processing them	✅ relaunch one c6in.16xlarge for ~10 min
3	peers_memory (not_yet_converted)	same shard; never reached convert phase	✅ same retry
1	scattered	error_write	⚠️ inspect ledger

For ML pretraining, filter WHERE conversion_status = 'ok' and you have 79,313.6 hours of lossless EEG.

Upstream data-quality patterns

Documented patterns from the per-dataset conversion ledgers. None of these are bugs in the canonical store — the stores either succeeded under the patterns below or were excluded.

Pattern	Example dataset	Auto-handled?
.vhdr missing MarkerFile= line	various	✅ patched in-place by the writer
.vhdr references a differently-named .eeg companion	various	✅ linked via the writer's internal .vhdr reference resolver
Numeric RDA_* BrainVision channels typed as misc	various	✅ blanket-promoted to eeg when all channels carry that prefix
EEGLAB .set references missing .fdt	ds002181, ds004306	❌ blocked — no upstream data
Corrupt .set metadata (sample-count mismatch, missing fields)	ds003343, ds003645, ds004166	❌ blocked
EEGLAB MAT v7.3 (HDF5-backed) .set files	several OpenNeuro releases	✅ handled via pymatreader
Epoched .set (no continuous signal)	ds004519, ds004771, ds005946	✅ writer concatenates epochs into a continuous stream + per-epoch annotations
Mixed-modality recordings (EEG + ECG + EMG)	various PhysioNet	✅ per-type unit map; non-µV-scalable channels recorded in dropped_channels

Corpus composition

The corpus mixes hand-curated anchor datasets, the OpenNeuro EEG bulk, and a foundation-model benchmark surface.

Anchor datasets

Dataset	Format	Channels	sfreq	License
HBN-EEG (Healthy Brain Network)	EEGLAB .set	129 (EGI HydroCel)	500 Hz	CC-BY-SA-4.0
PEERS Memory EEG (ds004395)	EDF / BDF	129 (EGI), 137, 144, 272 (BioSemi)	250–2048 Hz	CC0-1.0
TUH-EEG Corpus (TUEG family)	EDF	20–41 (clinical)	250 / 256 / 400 / 512 / 1000 Hz	TUH DUA
OpenNeuro EEG datasets (auto-ingested via manifest.openneuro_catalog)	mixed (.edf / .set / .vhdr / .bdf / .fif)	varies	varies	mostly CC0-1.0

Foundation-model benchmark / downstream eval

The datasets every EEG foundation model paper reports on, so a model trained on this corpus can be evaluated against the same tables LaBraM / CBraMod / REVE / DIVER-1 publish, without leaving the pipeline.

Dataset	Source	Recordings	Subjects	Hours	License
PhysioNet Sleep-EDFx — sleep cassette + telemetry	PhysioNet	197	197	3,849	ODC-By
PhysioNet HMC — Haaglanden Medisch Centrum sleep staging	PhysioNet	154	154	1,164	ODC-By
PhysioNet CAP — Cyclic Alternating Pattern sleep	PhysioNet	108	108	993	ODC-By
PhysioNet CHB-MIT — pediatric seizure	PhysioNet	665	23	962	ODC-By
PhysioNet Siena — adult scalp epilepsy	PhysioNet	38	14	130	ODC-By
PhysioNet MMI (EEGMMIDB) — motor imagery	PhysioNet	1,526	109	49	ODC-By
Mumtaz — MDD vs healthy controls	figshare 4244171	180	64	20	CC-BY-4.0
PhysioNet EEGMAT — mental arithmetic	PhysioNet	72	36	2	ODC-By
TUAB — TUH Abnormal	TUH NEDC	2,993	2,329	1,141	TUH DUA
TUAR — TUH Artifact	TUH NEDC	310	213	100	TUH DUA
TUEV — TUH Events	TUH NEDC	518	370	149	TUH DUA

PhysioNet datasets resolve via the PhysioNet AWS Open Data mirror (s3://physionet-open/); each is re-mirrored into the home bucket for same-region reads.

Licensing

The manifest itself (this Parquet bundle) is released under CC-BY-4.0. The underlying EEG signals are governed by their original licenses, and those licenses apply equally to the canonical .eegz stores derived from them.

Source family	Signal license	Access requirements
HBN-EEG	CC-BY-SA-4.0	Open; cite Shirazi et al. 2024
PEERS ds004395	CC0-1.0	Open
OpenNeuro bulk	almost all CC0-1.0	Open; cite the original upload per OpenNeuro convention
TUH-EEG family (TUEG, TUAB, TUAR, TUEV, TUSZ, TUEP, TUSL)	TUH DUA	Signed Data Use Agreement required via NEDC; the manifest only indexes metadata.
PhysioNet (MMI, Sleep-EDFx, CHB-MIT, Siena, CAP, EEGMAT, HMC)	ODC-By	Open; cite each PhysioNet record per Citation file
Mumtaz (figshare 4244171)	CC-BY-4.0	Open; cite Mumtaz et al. 2017

In particular, the manifest does not redistribute TUH signal data through the canonical store either. Those rows record file-level metadata only; canonical_uri points at a same-bucket store gated by the same DUA.

Companion code

The full pipeline that produces this corpus — schema definitions, S3 walkers, BIDS sidecar parser, EDF/BDF direct header parser, OpenNeuro mass-sync, canonical .eegz writer + reader + verifier, EC2 orchestrator — lives at ritivel/eegData.

# Convert one recording end-to-end (no S3 upload).
python -m manifest.eegz convert-one \
    --recording-id sub-001-ses-1 \
    --dataset-id   physionet_eegmat \
    --source-uri   s3://eeg-datasets-139156132535/raw/.../Subject00_2.edf \
    --source-format edf \
    --subject-id-in-dataset Subject00 \
    --no-upload

# Build the published manifest from the per-dataset conversion ledgers + canonical zarr.json.
python -m manifest.eegz_to_manifest \
    --input  s3://eeg-corpus-139156132535/manifest/latest/recordings.parquet \
    --output s3://eeg-corpus-139156132535/manifest/latest/recordings.parquet \
    --also-publish-latest

# Push the manifest to this Hugging Face dataset repo.
python -m manifest.publish_hf

Architectural notes and on-disk layout details are in docs/DESIGN.md; operator firefighting is in manifest/eegz/PLAYBOOK.md.

Citation

If this corpus accelerates your work, a one-line acknowledgment is appreciated and please also cite the underlying source datasets:

@misc{eeg_canonical_corpus,
  author = {Pavan Kalyan Tankala},
  title  = {EEG Canonical Corpus Index: A unified .eegz store over open-access EEG pretraining corpora},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/tankalapavankalyan/eeg-corpus-manifest}},
}