dpr/benchmark/README.md

# PhasePro DPR Benchmark

This directory is an exact copy of the PhasePro-derived DPR evaluation payload used by the local PhaseFlow workflow. It is an evaluation resource, not part of the PhaseFlow-DPR training data. The six payload files are kept byte-for-byte identical to `data/benchmarks/dpr_benchmark_phasepro/`; this README is release documentation only.

## Files

| File | Records | Purpose |
| --- | ---: | --- |
| `proteins.csv` | 121 proteins | Benchmark sequence records, identifiers, source membership, and split metadata |
| `regions.csv` | 140 regions | Tabular DPR annotations with both zero-based and one-based coordinates |
| `region_master.jsonl` | 139 regions | Canonical region records with evaluation metadata and tensor-coordinate fields |
| `region_spans.jsonl` | 119 proteins | Per-protein DPR and negative-span representation for evaluators |
| `source_records.csv` | 1,337 records | Source-level evidence and label provenance records |
| `source_map.csv` | 1,327 records | Mapping from source identifiers to benchmark protein identifiers |

Record counts exclude CSV headers. The apparent difference between `regions.csv`, `region_master.jsonl`, and `region_spans.jsonl` is expected because each file represents a different level of the annotation graph.

## Keys and coordinates

`protein_id` is the primary key across all files. `uniprot_id` and `sequence_hash` are useful secondary identifiers but should not replace `protein_id` in benchmark joins. `region_id` is the key for region-level records.

`regions.csv` carries both conventions: `start`/`end` are zero-based half-open coordinates (`[start, end)`), while `start_1based`/`end_1based` describe the same interval in one-based inclusive form. `region_master.jsonl` explicitly marks `coordinate_system` as `1-based inclusive`; its `tensor_start` is zero-based and `tensor_end` is the matching exclusive endpoint. Preserve these conventions when evaluating predicted residue scores.

## Evaluation boundary

Keep all files in this directory outside the training sampler, hyperparameter search, threshold selection, and checkpoint choice. Training data are under `../data/`; their `protein_id` values are disjoint from `proteins.csv`. If a downstream workflow requires a different split or a transformed representation, retain the original files unchanged and write derived artifacts elsewhere.

## Provenance and attribution

The source fields retain the provenance assembled by the local benchmark workflow. Please cite the original PhaSePro/PhasePro resource when using or discussing these benchmark-derived annotations.