# VariantHound Hugging Face data contract

## Package layout

| Path | Purpose |
|---|---|
| `data/benchmark_cases.jsonl` | One validated canine benchmark case per line, sorted by case ID |
| `data/source_manifest.json` | Upstream source access, pinning, redistribution, and attribution policy |
| `data/ensembl-compara-116.json` | Immutable Ensembl 116 artifact checksums, assemblies, and expected row counts |
| `data/hpo-2026-09-01.json` | Immutable HPO asset IDs, checksums, and expected normalized counts |
| `data/monarch-kg-2026-09-02.json` | Dated Monarch/HGNC artifacts and exact normalized counts |
| `data/mgi-2026-09-12.json` | Retrieval-dated MGI/MP artifacts, conflict policy, and exact counts |
| `data/omia-2026-09-12.json` | Retrieval-dated OMIA dump and official gene-crosswalk checksums |
| `data/phenio-2026-09-01.json` | Immutable PHENIO relation-graph assets and exact compact-closure counts |
| `data/orthology_manifest.json` | Release-qualified dog-human-mouse mapping policy or pinned mapping snapshot |
| `data/hpo/snapshot.json` | HPO release metadata and normalized-file integrity records in full releases |
| `data/hpo/ontology_terms.jsonl` | HPO term hierarchy, synonyms, alternate IDs, and obsolescence |
| `data/hpo/gene_phenotypes.jsonl` | NCBI Gene-to-HPO annotations with disease and frequency context |
| `data/hpo/disease_annotations.jsonl` | Disease-to-HPO annotations with references, evidence, and negation |
| `data/monarch/snapshot.json` | Monarch release, deduplication, provenance, and output integrity records |
| `data/monarch/human_gene_disease.jsonl` | Stable-ID human gene-to-MONDO evidence |
| `data/monarch/mouse_gene_phenotype.jsonl` | MGI gene-to-MP evidence with primary-source attribution |
| `data/mgi/snapshot.json` | Direct MGI provenance, identifier policies, and output integrity records |
| `data/mgi/mp_terms.jsonl` | Mammalian Phenotype hierarchy and identifier status |
| `data/mgi/gene_crosswalk.jsonl` | Phenotype-bearing MGI markers with reconciled NCBI/Ensembl IDs |
| `data/mgi/genotype_phenotypes.jsonl` | Direct genotype-to-MP evidence and rank-eligibility state |
| `data/omia/snapshot.json` | Canine OMIA provenance, masking and output-integrity contract |
| `data/omia/genes.jsonl` | OMIA internal-to-NCBI Gene resolution for canine genes |
| `data/omia/phenes.jsonl` | Canine phenes, clinical text, inheritance, breeds and ontology links |
| `data/omia/gene_phene_evidence.jsonl` | Stable-ID canine gene–phene evidence with discovery mask IDs |
| `data/omia/causal_variants.jsonl` | Variants linked to canine single-locus traits with assemblies and publications |
| `data/omia/publications.jsonl` | Publications referenced by canine OMIA phenes or variants |
| `data/phenio/snapshot.json` | PHENIO release, closure semantics, prefix policy, and output integrity |
| `data/phenio/phenotype_closure.tsv.gz` | Entailed reflexive HP/MP/uPheno subclass closure used for semantic similarity |
| `data/phenio/upstream_versions.tsv` | Versions of all ontologies integrated into the pinned PHENIO release |
| `data/browser/manifest.json` | Content-addressed browser collection map, source fingerprints, counts, and limitations |
| `data/browser/{collection}/*` | Sixteen deterministic gzip buckets per terms, closure, gene-profile, and orthology collection |
| `data/ml-readiness-policy.json` | Versioned independent-sample and learned-model promotion thresholds |
| `schemas/*.schema.json` | Machine-readable contracts used by the source repository |
| `docs/*.md` | Data provenance, leakage-safe evaluation, source-ingestion, and ranking guidance |
| `release-manifest.json` | Dataset version, source commit, file sizes, and SHA-256 digests |

The release builder uses an explicit source-to-destination allowlist. Files under `data/raw`, `data/processed`, browser build directories, credentials, VCF/BCF files, and complete third-party database exports are never copied implicitly. The complete OMIA SQL dump is excluded; only its verified canine transform can be selected.

## Benchmark case fields

Each line in `data/benchmark_cases.jsonl` contains:

- `id`, `label`, `species`, `breed`, and `expected_gene`;
- `disease_group_id`, the indivisible split and resampling unit;
- `phenotypes.present` and `phenotypes.excluded` ontology identifiers;
- `publication_year`;
- `evaluation.discovery_mask_evidence_ids` and `evaluation.discovery_mask_disease_ids`;
- `evaluation.temporal_status` and an exclusion reason when applicable;
- a literature `reference` with label and URL.

Cases are benchmark inputs, not independent training rows. Disease-level grouping and temporal availability must be preserved in any downstream evaluation.

## Release integrity

`release-manifest.json` is valid only when:

1. every listed payload path exists and no unlisted file is present;
2. every byte count and lowercase SHA-256 digest matches;
3. the dataset version and source Git commit are recorded;
4. the package was built after the repository data validator passed.

The upload script rechecks all four conditions before contacting Hugging Face.