# bciciv-2a — public training data

Public-side data for BrainPilotBench task
[`tasks/bciciv-2a`](https://github.com/NeuroAIHub/BrainPilotBench/tree/main/tasks/bciciv-2a).

Held-out session-E evaluation GDF files and their true labels live in the
gated companion dataset
[`BrainPilot-Bench/Tasks-Data-Private`](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Private)
and never appear here.

## Files

| File | Contents | Size |
|---|---|---|
| `public.tar.zst` | `gdf/A0[1-9]T.gdf` (9 subjects × session-T training GDF), `manifests/{subjects.json,protocol.json}`, `get_subject_data.py`, `label_mapping.json`, `README.md` | ~ 194 MB |

`public.tar.zst` is pinned by sha256 in the task's `data.lock`; the agent's
`env/setup-agent.sh` untars it into the workspace `public_data/`.

## Task summary

BCI Competition IV Dataset 2a — 4-class motor imagery decoding (left hand /
right hand / feet / tongue). 9 subjects × 22 EEG channels × 288 trials per
session. The agent submits a PyTorch `MIAgentModel` module; the benchmark
scorer trains one model per subject on session T (230/58 stratified train/val
split, fixed seed 42) and evaluates on session E (288-trial blind test).

See the task README in the main repo for the full contract.

## Provenance & license

GDF files come from the BCI Competition IV Dataset 2a (Graz). License
Apache-2.0 (see repo `LICENSE`). See
<https://www.bbci.de/competition/iv/> for the original release.