# Data Dictionary

## Training data

| File | Unit of observation | Required keys | Description |
| --- | --- | --- | --- |
| `data/proteins.parquet` | protein sequence | `protein_id`, `sequence_sha256` | Unique amino-acid sequence records and sequence lengths. |
| `data/training_units.parquet` | protein-stage assignment | `protein_id`, `sequence_sha256` | Indicates whether a protein participates in base training, region refinement, or both. |
| `data/base_training_schedule.parquet` | scheduled training draw | `protein_id`, `sequence_sha256`, `update`, `rank`, `slot` | Deterministic base-training ordering, sampling tier, source, and label metadata. |
| `data/region_supervision.parquet` | region-supervision record | `protein_id`, `sequence_sha256`, `supervision_id` | Final DPR region targets, evidence provenance, loss permissions, and region boundaries. |

Join all training tables to `data/proteins.parquet` on both `protein_id` and `sequence_sha256`. In `region_supervision.parquet`, `region_start` and `region_end` are sequence-index boundaries; consult the published PhaseFlow workflow before converting them to model tensors.

## Evaluation benchmark

`benchmark/` is evaluation-only and is documented in [`../benchmark/README.md`](../benchmark/README.md). The benchmark’s `proteins.csv` has no `protein_id` overlap with `data/training_units.parquet`. Never include benchmark files in a sampler or use them for model selection.

## Configuration and integrity metadata

`configs/dataset_contract.yaml` defines the package contract and `configs/training.yaml` records public training settings. `metadata/feature_reproduction.md` describes feature reconstruction boundaries. `metadata/file_inventory.csv` lists every package file other than itself with its byte size and SHA-256 digest.