# sleep-edf — public training data

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

Held-out subject 16-19 evaluation recordings live in the gated companion
dataset
[`BrainPilot-Bench/Tasks-Data-Private`](https://huggingface.co/datasets/BrainPilot-Bench/Tasks-Data-Private)
and never appear as "test" here.

## Files

| File | Contents | Size |
|---|---|---|
| `public-main.tar.zst` | `sleep-cassette/SC4[0-1][0-5][12]E[0-9]-PSG.edf` + `-Hypnogram.edf` (subjects **0–15**, 32 files), `manifests/{subjects.json,default_split.json,loso_folds.json}`, `get_fold_data.py`, `label_mapping.json`, `README.md` | ~ 934 MB |
| `loso-supplement.tar.zst` | Additional PSG+Hypnogram for subjects **16-19** (16 files) — **only** for agents that want to run the 20-fold LOSO protocol over all 20 subjects. Do NOT use as test data under the default split. | ~ 235 MB |

`public-main.tar.zst` is the default fetch and pinned in the task's
`data.lock`. `loso-supplement.tar.zst` is optional and only needed for LOSO
runs; the task's data.lock exposes it as an opt-in dataset (`bp-bench fetch
sleep-edf --extra loso`).

## Task summary

Sleep-EDF Expanded — Sleep Cassette. 5-class sleep stage classification
(Wake / N1 / N2 / N3 / REM) on 30-second Fpz-Cz EEG epochs. 20 subjects
split by subject boundary: train subjects 0-13, validation 14-15, test
16-19 (private). The agent submits a PyTorch `SleepAgentModel` module; the
benchmark scorer trains one model on subjects 0-13, uses 14-15 for early
stopping, and blind-tests on 16-19.

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

## Provenance & license

Sleep-EDF Database Expanded v1.0.0 — <https://physionet.org/content/sleep-edfx/1.0.0/>.
Files are re-hosted under the same open license as PhysioNet.