# Dataset files

This HF payload includes all 3,179 original protein graphs, unlike the separate GitHub code release. Raw numerical inputs, feature caches, folds, targets, and saved outputs are preserved. No scientific experiments were run during preparation.

- raw/: original aligned RMSD, PoseBusters-validity, and runtime matrices.
- targets/: acquisition target serializations and the explicitly reconstructed native evaluation target; their definitions differ.
- splits/: original five-fold assignments and the independently recovered membership cross-check.
- descriptors/: four original physical descriptors per complex.
- features/: original 256-dimensional projected cache, verified IDs, and original ligand embeddings.
- external/protein_graphs_esmc_600m/: all 3,179 original serialized protein graph files.
- viewer/: deterministic cellwise reshaping and a path/size/hash graph index for HF configuration support.
- provenance/: unchanged original data-source and validation records. Historical source states retain their original context.

The complete schema, units, source-quality limits, target definitions, and original preprocessing are in ../docs/data_card.md. The root dataset card defines nine configurations with split full; use the original test_fold/outer_test_fold assignments for scientific evaluation rather than randomly splitting cells.

Run `python3 scripts/validate_data.py --require-complete` and `python3 scripts/verify_hf_package.py` from the dataset root. New reshaping/index tables do not replace the original data. Author-owned contributions use CC BY 4.0; separate upstream terms apply to third-party components (../LICENSE).