GraphRareBench Dataset

This release is the single frozen GraphRareBench benchmark described in the AAAI 2027 submission. It contains 2,365 ontology-derived rare-disease ranking cases and 18,093 labeled target-confounder pairs.

Layout

dataset/
  cases/graphrarebench_cases.jsonl
  sidecars/evidence_bundle.jsonl
  sidecars/graph_path_alignment.jsonl
  metadata/manifest.json
  metadata/release_summary.json
  metadata/checksums.sha256
  schema/case_public_schema.json
  schema/evidence_bundle_public_schema.json
  provenance/source_provenance.csv

Case File

cases/graphrarebench_cases.jsonl is the benchmark anchor. Each row contains:

• case_id, split, and partition_strategy;
• query_phenotypes: the coarsened HPO query visible to evaluated methods;
• candidate_pools.full: the closed full candidate set;
• candidate_pools.hard: the target plus graph-defined hard confounders;
• target_disease and hard_confounders: evaluator-only labels for scoring;
• evidence_bundle_id and graph_path_alignment_refs for source-linked audit.

The public split is gene-component-aware: 1,892 train, 236 dev, and 237 test cases. Test cases have a median of 69 full-pool candidates and a median of 5 hard confounders.

Evaluation

Full-pool ranking uses candidate_pools.full. The tool-mediated hard-pool audit uses candidate_pools.hard. Evaluated systems should receive only the query, candidate IDs/names, and permitted evidence interface; target labels, hard-confounder labels, and mechanism annotations are evaluator-only.