zifeng-ai/rare-disease-diagnosis

Rare Disease Diagnosis v1

Materialized task data for the deeprare-curated-phenotype/v1 agent benchmark in T0-RSI/ai4sci-tasks. The files preserve the materializer's host directory layout. This is a frozen scientific case selection served from the current repository main; later Hub commits may update packaging or documentation.

• —public/development/: labeled fit/tune examples for the agent.
• —public/verifier/: phenotype-only evaluation inputs.
• —public/reference/: read-only HPO/Orphadata sources and derived search indexes.
• —private/: evaluation labels and provenance, mounted only by the isolated grader during benchmark runs. These files are publicly downloadable from this dataset, so the held-out answers are no longer secret.

Download the complete repository on the host. Do not mount private/ into the agent or verifier containers. The root release-manifest.json checks every data file, including references; the existing case and reference manifests retain their source provenance and scientific hashes.

This collection contains 1,268 cases from RAMEDIS, MME, HMS, LIRICAL, and MyGene2; 740 cases are held out for evaluation. See NOTICE.md and the reference source_manifest.json for source credits, versions, and terms.

This dataset uses the Human Phenotype Ontology (HPO v2025-05-06), developed by the Human Phenotype Ontology Consortium. Cite Köhler et al., Nucleic Acids Research 42(D1), D966–D974 (2014), doi:10.1093/nar/gkt1026. The original HPO ontology and annotation files are byte-preserved. It also uses Orphadata Science's July 2025 phenotypes file under CC BY 4.0; credit Orphanet/Orphadata. The source cases were publicly released by RareBench and MyGene2/SHEPHERD. Public evaluation labels make this release unsuitable as a secret-label benchmark. The upstream licenses and attribution requirements still apply.