data/sample/README.md

# Sample dataset (in-repo)

This directory ships a single pre-built target — **`cam` (Calmodulin)** — so users
can run a smoke test of the training and evaluation pipeline without first
downloading the full multi-target dataset (~10 GB on Zenodo) or rebuilding
from raw PDB files (~30 min per target).

## Contents

```
sample/
└── cam/
    ├── train.pkl   # 84 paired holo/apo complex graphs   (~24 MB)
    ├── val.pkl     # 12 validation graphs                (~1.3 MB)
    └── test.pkl    # 96 held-out evaluation graphs       (~25 MB)
```

Each pickle is a list of dicts produced by `code/data/build_dataset.py`.
Splits follow the family-stratified scheme used in the paper
(equivalent to `data/processed_familysplit/cam/` train+val and
`data/processed_familysplit_v5/cam/test.pkl` in the source tree).

## Smoke test (1-epoch end-to-end)

```bash
# Train both phases for 1 epoch
python code/scripts/train.py \
    --target cam \
    --phase both \
    --data_dir data/sample \
    --checkpoint_dir checkpoints_smoke \
    --epochs 1 \
    --no_wandb

# Evaluate
python code/scripts/evaluate.py \
    --target cam \
    --checkpoint checkpoints_smoke/best_phase2.pt \
    --data_dir data/sample \
    --outdir eval_smoke
```

Expected runtime: ~1 minute on a single GPU.

## Want more data?

- All 12 paper targets, pre-built: see `data/DOWNLOAD.md` for the Zenodo link.
- Build from raw PDBs locally: `scripts/build_data.sh paper12`.
- Per-target PDB lists and chain mappings: `data/target_lists/*.txt` (68 targets).