license: cc0-1.0
tags:
- eeg
- bci
- brain-computer-interface
pretty_name: AI4Sci BCI decoding — public tracks (AD65, OpenBMI, Longitudinal)
AI4Sci brain-computer-interface decoding: public tracks
Prepared data for three of the five tracks of the brain-computer-interface-decoding task in
T0-RSI/ai4sci-tasks and
sunnweiwei/science-speedrun. Each was built by
the task's environment/data/materialize.py from pinned public sources:
| track | source | classes | shape (channels, seconds, 200 Hz) | train / val / test windows |
|---|---|---|---|---|
ad65 |
OpenNeuro ds004504 derivatives (CC0) | 3: Alzheimer's, frontotemporal dementia, control | 19 × 10 × 200 | 3,774 / 1,343 / 1,376 |
openbmi |
GigaDB 100542 motor-imagery sessions (CC0) | 2: left, right | 62 × 4 × 200 | 12,800 / 4,400 / 4,400 |
longitudinal |
OpenNeuro ds007176 (CC0) | 2: eyes closed, open | 58 × 2 × 200 | 28,523 / 8,418 / 10,028 |
Subjects are split with seed 3407 (test 20%, then validation 25% of the rest; AD65 stratified by diagnosis); every session stays with its subject. Arrays are float32 microvolts in NumPy shards.
runtime/development/<track>/{train,val}/ manifest.json (id, label, subject, session, path, index), shards
runtime/private/<track>/public/ test inputs: manifest without labels, shards in random order
runtime/private/<track>/truth.jsonl test labels, subjects and sessions (verifier only)
records/<track>/ subject splits, preprocessing statistics, build receipt
release-manifest.json size and SHA-256 of every file
The FACED and TUEV tracks are not redistributable and come from a separate controlled-access
release. The task's environment/data/materialize.sh release ROOT installs all five tracks and
verifies every file against release-manifest.json. Only runtime/development may be mounted
into an agent's workspace; runtime/private is for the verifier.
Citations: Miltiadous et al. (2023), Data 8(6), 95 (ds004504); Lee et al. (2019), GigaScience 8(5), giz002 (OpenBMI); OpenNeuro ds007176.