# Initial 2,000-protein dataset

This is the canonical local root for the first complete router dataset: 1,000
single-dominant structured-state proteins (label `0`) and 1,000 dynamic or
heterogeneous-state proteins (label `1`). The fixed split is 1,400 train, 300
validation, and 300 test proteins.

Place Colab's completed ESMFold result files (`<sequence_sha256>.npz`) in
`esmfold_results/`. Then import them with:

```bash
uv run python scripts/esmfold_dataset.py import
```

The importer validates every NPZ against `esmfold_input_manifest.csv` and
writes the reusable package-native bundles into `embeddings/esmfold_v1/`. It
also writes `embeddings/esmfold_v1/manifest.parquet`, the index consumed by the
training commands. Do not train directly from `esmfold_results/`; it is only a
transfer/staging location and can be deleted after a successful import.

Key files:

- `catalog.parquet` and `catalog.csv`: canonical labels, sequences, metadata,
  and fixed split assignments.
- `catalog_esmfold_eligible.parquet` and `.csv`: the 1,988 rows with verified
  ESMFold bundles; use these for embedding-only or combined model training.
- `esmfold_input_manifest.csv`: exact sequence hashes and model input contract.
- `embeddings/esmfold_v1/manifest.parquet`: verified embedding-bundle index.
- `embeddings/esmfold_v1/excluded.csv`: the 12 sequences outside ESMFold's
  1,022-residue limit.

For CPU/MPS training, pass `--dataset data/initial_2000_dataset/catalog.parquet`
and `--bundle-manifest data/initial_2000_dataset/embeddings/esmfold_v1/manifest.parquet`.