The Dataset Viewer has been disabled on this dataset.

Amber — an open EEG corpus

6,046 hours of EEG from 7,275 recordings, 1,864 subjects and 164 public datasets, all converted to a single self-describing container format so they can be read with one line of code regardless of where they came from.

Recording hours 6,046.4
EEG channel-hours 37,487
Recordings 7,275
Files 7,523
Subjects 1,864
Source datasets 164
Annotated events 3,612,642
Size 60.5 GB
Bit-exact files 3,620 / 7,523 (48%)

What the subjects were doing

The corpus is assembled to favour EEG of people doing something — reading, listening, remembering, deciding, controlling an interface — over recordings of people lying still. Labels are derived from dataset titles and task entities by tools/content_tiers.py, so they are a guide rather than a curated ground truth.

Content Hours Recordings Datasets
active task 295.3 3,434 61
clinical 164.8 83 5
resting / unlabelled 336.8 3,382 92
sleep 5,249.5 376 6

Quick start

The reader needs only numpy and zstandard; the heavier converters (pyedflib, mne) are imported lazily and are not required to read:

pip install numpy zstandard huggingface_hub
from huggingface_hub import snapshot_download
import eegx  # reader shipped in this repo under eegx/

p = snapshot_download('amongglue/amber', repo_type='dataset',
                      allow_patterns='data/physionet/eegmat/*')
r = eegx.open(f'{p}/data/physionet/eegmat/Subject00_1.eegx')

x = r.read()                 # (n_channels, n_samples) float32, physical units
x = r.read(picks=['eeg'])    # EEG channels only
print(r.sfreq, r.duration, r.channel_names, r.events)

Every file is one (n_channels, n_samples) channel-major array. That is the format's single rigidity guarantee: whatever the montage, the task or the original vendor format, this shape always holds, so an unknown recording can still be enumerated and converted without special-casing. See SPEC.md for the container layout.

Layout

data/<archive>/<dataset>/*.eegx    signals, one file per recording
data/<archive>/<dataset>/README.md source, licence, hours, channels
catalog.jsonl                      one row per file: id, hours, channels, rates
SPEC.md                            the container format
eegx/                              reference reader (pure Python + numpy)

catalog.jsonl is the index to work from — it lets you filter by sample rate, channel count or duration without opening any signal file.

Contribution by source

Other — 6,046.4 h across 164 datasets

Dataset Source Hours Rec Subj Rate (Hz) Ch Licence
sleep-edfx 3,849.0 197 100 1-100 1-4 ODC-BY
hmc-sleep-staging 1,144.2 151 151 256 8 ODC-BY
ucddb 173.4 25 25 8-128 3-8 ODC-BY
siena-scalp-eeg 141.0 41 14 512 35-49 ODC-BY
ds006366 72.0 1 1 128 2 ODC-BY
eegmmidb 48.5 1,526 109 128-160 64 ODC-BY
ds003190 39.8 384 19 256 9-10 ODC-BY
ds004369 34.7 38 38 500 7 ODC-BY
ds007720 31.7 371 5 300 7-20 ODC-BY
ds007554 29.9 498 30 250-250.003 33 ODC-BY
ds005342 28.8 28 28 250 17 ODC-BY
moabb_BNCI2014_004 26.3 45 9 250 7 ODC-BY
chbmit 21.7 22 1 256 23 ODC-BY
ds007058 16.9 178 10 200 63 ODC-BY
moabb_Kaya2018 15.7 17 7 200 19 ODC-BY
moabb_BNCI2015_004 13.8 18 9 256 31 ODC-BY
moabb_BNCI2015_012 13.6 20 10 250 65 ODC-BY
ds005628 13.3 140 102 250 8 ODC-BY
ds005340 11.7 5 5 10000 2 ODC-BY
moabb_BNCI2014_001 11.6 108 9 250 26 ODC-BY
ds004324 11.5 16 16 500 38 ODC-BY
ds006260 10.9 179 72 256 32 ODC-BY
ds004860 10.6 20 20 512-2048 41 ODC-BY
ds007541 10.3 960 32 1000 8 ODC-BY
shhpsgdb 9.0 1 1 250 11 ODC-BY
moabb_BNCI2025_002 8.4 27 3 200 64 ODC-BY
ds004661 8.2 14 14 128 64 ODC-BY
osf_v9xsw 7.7 5 5 500 63 ODC-BY
ds007864 7.0 32 32 200 19-21 ODC-BY
ds004577 6.9 31 31 200 19-24 ODC-BY
ds007558 6.9 35 35 200 19-21 ODC-BY
ds007666 6.9 34 34 200 19 ODC-BY
ds007667 6.9 34 34 200 19 ODC-BY
ds007865 6.8 33 23 200 19-21 ODC-BY
ds005262 6.6 49 12 250 8 ODC-BY
ds004840 6.4 32 9 256-1024 9-10 ODC-BY
ds006839 6.3 29 29 1000 29 ODC-BY
ds006394 6.2 21 21 125 16 ODC-BY
moabb_BNCI2015_013 6.1 120 6 512 65 ODC-BY
ltrsvp 5.9 62 11 2048 9 ODC-BY
ds004752 5.3 48 14 200-4096 8-23 ODC-BY
ds003194 5.1 21 15 200 19-20 ODC-BY
moabb_Nguyen2017_V 4.8 8 8 256 65 ODC-BY
ds002691 4.5 13 13 250 32 ODC-BY
ds005383 4.2 122 30 200 31 ODC-BY
moabb_Dreyer2023C 4.2 36 6 512 32 ODC-BY
moabb_BNCI2003_004 4.0 5 5 100 118 ODC-BY
moabb_BNCI2019_001 3.8 45 5 256 64 ODC-BY
hbn_contrast-change-detection 3.7 34 13 250-500 129 ODC-BY
moabb_Mainsah2025_K 3.6 128 5 256 17 ODC-BY
moabb_Nguyen2017_S 3.6 6 6 256 65 ODC-BY
figshare_32676945 3.5 2 2 1000 33 ODC-BY
ds004771 3.5 20 20 256 34 ODC-BY
ds002720 3.2 30 18 1000 19 ODC-BY
ds005420 3.2 43 37 500 20 ODC-BY
moabb_BNCI2014_008 3.0 8 8 256 10 ODC-BY
moabb_EPFLP300 3.0 192 8 2048 35 ODC-BY
ds003195 2.7 11 10 200 19 ODC-BY
osf_x8wfv 2.7 3 3 512 70 ODC-BY
ds006593 2.6 10 10 300 19 ODC-BY
eegmat 2.4 72 36 500 21 ODC-BY
moabb_Nguyen2017_L 2.4 6 6 256 65 ODC-BY
ds007169 2.3 8 8 250 24 ODC-BY
erpbci 2.3 245 12 2048 70 ODC-BY
moabb_Nguyen2017_SL 2.2 6 6 256 65 ODC-BY
moabb_Wang2021Combined 2.1 8 2 1000 33 ODC-BY
moabb_Nakanishi2015 2.1 9 9 256 9 ODC-BY
hbn_surround-suppression 2.0 20 11 250-500 129 ODC-BY
ds004368 1.9 9 9 128 63 ODC-BY
ds006018 1.9 8 8 500 31 ODC-BY
ds007823 1.9 8 8 200 22 ODC-BY
ds008082 1.9 8 8 200 22 ODC-BY
ds003574 1.9 1 1 500 69 ODC-BY
ds005863 1.8 9 8 500 31 ODC-BY
ds005274 1.7 22 22 500 6 ODC-BY
ds007262 1.7 7 7 250 24 ODC-BY
ds004588 1.7 13 13 300 24 ODC-BY
figshare_24631257 1.7 1 1 1024 145 ODC-BY
moabb_GrosseWentrup2009 1.7 2 2 500 129 ODC-BY
zenodo_13626922 1.6 16 16 128 13 ODC-BY
moabb_BNCI2014_009 1.6 30 10 256 18 ODC-BY
ds007822 1.6 32 32 300 19 ODC-BY
figshare_29972935 1.6 8 8 500 35 ODC-BY
ds007788 1.3 29 7 100 64 ODC-BY
ds004022 1.1 5 5 500 18 ODC-BY
moabb_AlexMI 1.1 8 8 512 17 ODC-BY
ds003800 1.0 20 13 250 19 ODC-BY
ds008062 1.0 30 30 128 14 ODC-BY
moabb_Beetl2021_A 1.0 6 3 500 63 ODC-BY
ds004015 1.0 1 1 500 18 ODC-BY
ds003810 1.0 8 8 125 15 ODC-BY
ds002218 0.9 1 1 256 32 ODC-BY
ds006465 0.9 3 3 1000 127 ODC-BY
moabb_BNCI2015_003 0.9 20 10 256 10 ODC-BY
ds008083 0.9 1 1 1000 31 ODC-BY
moabb_Mainsah2025_E 0.9 33 3 256 17 ODC-BY
ds007006 0.9 7 7 256 64 ODC-BY
hbn_sequence-learning 0.9 9 9 250-500 129 ODC-BY
ds006036 0.8 8 8 500 19 ODC-BY
moabb_Mainsah2025_N 0.8 20 1 256 17 ODC-BY
ds007629 0.8 4 1 256 64 ODC-BY
ds001787 0.8 1 1 256 80 ODC-BY
ds002721 0.8 9 9 1000 19 ODC-BY
moabb_Beetl2021_B 0.7 4 2 200 32 ODC-BY
ds003801 0.7 1 1 250 24 ODC-BY
figshare_30227503 0.7 1 1 1000 33 ODC-BY
figshare_31939692 0.6 1 1 1000 64 ODC-BY
ds002778 0.6 11 11 512 41 ODC-BY
ds004018 0.6 1 1 1000 63 ODC-BY
ds004541 0.6 1 1 1000 69 ODC-BY
figshare_25117430 0.6 1 1 500 35 ODC-BY
hbn_film-watching 0.6 12 4 250-500 129 ODC-BY
ds004849 0.5 1 1 128 64 ODC-BY
ds004850 0.5 1 1 128 64 ODC-BY
ds004852 0.5 1 1 128 64 ODC-BY
ds004853 0.5 1 1 128 64 ODC-BY
ds004854 0.5 1 1 128 64 ODC-BY
ds007406 0.5 10 10 256 14 ODC-BY
ds004785 0.5 9 9 500 32 ODC-BY
ds007955 0.5 9 9 250 8 ODC-BY
hbn_processing-speed 0.4 8 8 250-500 129 ODC-BY
ds007813 0.4 1 1 500 63 ODC-BY
ds004196 0.4 1 1 512 73 ODC-BY
ds004515 0.3 1 1 500 66 ODC-BY
ds007591 0.3 1 1 256 139 ODC-BY
ds003753 0.3 1 1 500 66 ODC-BY
ds004595 0.3 1 1 500 66 ODC-BY
ds004745 0.3 1 1 1000 8 ODC-BY
ds004347 0.3 1 1 512 73 ODC-BY
figshare_28389305 0.3 1 1 1000 56 ODC-BY
ds007688 0.3 2 2 250 73 ODC-BY
ds008446 0.2 1 1 512 65 ODC-BY
ds003039 0.2 1 1 500 67 ODC-BY
ds003517 0.2 1 1 500 65 ODC-BY
figshare_30740115 0.2 1 1 250 16 ODC-BY
figshare_31315511 0.2 1 1 250 16 ODC-BY
ds006525 0.2 1 1 250 128 ODC-BY
ds003490 0.2 1 1 500 67 ODC-BY
ds003751 0.2 1 1 250 132 ODC-BY
ds007172 0.1 1 1 500 32 ODC-BY
ds005841 0.1 1 1 512 73 ODC-BY
ds006126 0.1 1 1 5000 3 ODC-BY
ds005284 0.1 1 1 1024 65 ODC-BY
ds005305 0.1 1 1 512 64 ODC-BY
ds005872 0.1 1 1 500 129 ODC-BY
ds007338 0.1 1 1 500 129 ODC-BY
ds006768 0.1 1 1 1000 65 ODC-BY
ds003775 0.1 1 1 1024 64 ODC-BY
ds004381 0.1 2 2 20000 4 ODC-BY
hbn_other 0.1 1 1 500 129 ODC-BY
ds007222 0.1 1 1 500 266 ODC-BY
figshare_30162868 0.1 1 1 512 1 ODC-BY
hbn 0.1 1 1 500 129 ODC-BY
ds003805 0.0 1 1 500 19 ODC-BY
figshare_29664056 0.0 1 1 1000 62 ODC-BY
ds005028 0.0 40 11 40272-106312 32 ODC-BY
figshare_29528219 0.0 1 1 200 32 ODC-BY
figshare_30003127 0.0 1 1 200 32 ODC-BY
figshare_31081684 0.0 1 1 512 64 ODC-BY
ds002893 0.0 1 1 250.029 47 ODC-BY
ds003380 0.0 0 1 2000 16 ODC-BY
ds004117 0.0 1 1 250 47 ODC-BY
ds005048 0.0 16 16 250 43 ODC-BY
figshare_29170289 0.0 1 1 128 32 ODC-BY

How hours are counted

Recording hours is wall-clock time, counted once per source recording. Recordings that mix sample rates are split into one file per rate group, so a single night of sleep may be several files; only files carrying at least one EEG channel count, and parts of the same recording are never added together. Channel-hours multiplies duration by EEG channel count, which is the number that matters for training-data volume.

Fidelity

EDF/BDF sources are read at the digital level and stored as the same integers with the original per-channel affine map, so the conversion is bit-exact and reversible — verifiable with tools/verify_lossless.py in the source repository, which re-downloads the original and compares samples. Sources that only expose floating-point data (EEGLAB .set, some BrainVision) are requantised to int16 against a per-channel full-scale range; those files are marked lossless: false in their header and in catalog.jsonl.

Provenance and licences

Every file records where it came from — source URL, SHA-256 of the original, licence, and conversion time:

eegx.read_header(path)['provenance']

All included datasets are CC0 or ODC-BY 1.0, both of which permit redistribution. ODC-BY requires attribution: the per-dataset README.md names the original source, and those citations should travel with any derived work. This repository is a format conversion, not new data — please cite the original authors, not just this corpus.

Downloads last month
247