# ICASSP 2027 EEG-sAAD Challenge workspace

## Layout

- `archives/datasets/` — verified, original dataset ZIP archives and `SHA256.txt`.
- `data/cnd/` — extracted extended-CND training datasets; this is the input for the official MATLAB baseline and the converter.
- `data/superhuge/` — reserved output for CND-to-SuperHuge conversion, used by Python DNN baselines.
- `baselines/matlab_linear/` — official Linear stimulus-reconstruction and causal-HMM baseline.
- `baselines/matlab_scoring/` — official prediction scoring code.
- `baselines/python/SuperHugeAAD/` — Python DNN baselines (VLAAI and SSMamba), checked out on `python-package-working`.
- `tools/convert_cnd_to_superhuge.py` — supplied CND-to-SuperHuge conversion tool.
- `experiments/` — local training runs, logs and checkpoints.
- `submissions/` — prediction files and submission-ready artifacts; do not place raw data here.
- `docs/reference/` — source documentation.

## Python environment

```bash
cd /root/disk_3tb/EEG-sAAD-Grand-Challenge-Training-Dataset
source .venv/bin/activate
```

The environment uses Python 3.12.13, PyTorch 2.6.0+cu124, CUDA-enabled Mamba SSM, Lightning and the SuperHuge package. It has been smoke-tested on the local Tesla V100.

## Data workflow

1. Preserve the original archives under `archives/datasets/`.
2. Use the extracted CND data in `data/cnd/` for the MATLAB Linear SR baseline.
3. When required, convert CND data into `data/superhuge/` for Python DNN experiments:

```bash
.venv/bin/python tools/convert_cnd_to_superhuge.py \
  --source-root data/cnd \
  --out-root data/superhuge
```

The conversion creates a large derived dataset and is intentionally not run automatically.
