GraphRareBench Dataset Card
Summary
GraphRareBench is a provenance-preserving benchmark for phenotype-driven rare-disease ranking. The frozen release contains 2,365 ontology-derived cases, 18,093 target-confounder pairs, coarsened HPO queries, fixed candidate pools, graph-defined hard confounders, and source-linked evidence records.
Intended Use
Use the release to evaluate closed candidate-set rare-disease ranking methods, target-over-confounder discrimination, mechanism-specific failures, and tool-mediated evidence-access behavior.
Splits
- Train: 1,892 cases
- Dev: 236 cases
- Test: 237 cases
Diseases connected through the same causal-gene component are assigned to the same partition when gene annotations are available.
Metrics
The accompanying code computes MRR, Hit@1/5/10, ToC_case, ToC_pair, AnyConfAbove, mechanism-sliced ToC_pair, case-level bootstrap confidence intervals, paired bootstrap contrasts, and deterministic tool-trace audit metrics.