Same items as iit-patna-cse-ai/GenBench_test (regenerated after adding a confidence gate (>= 0.5) to GENE_ASSOCIATED_WITH_DISEASE-derived answers in counterfactual and disease_reasoning (coding_variant Tier D) task types -- a real check found 72% of disease edges previously used as answers had confidence < 0.5, mostly single-source OpenTargets/KEGG associations with no corroborating evidence), with three added columns from an independent per-item verification pass:

is_valid (0/1): whether the item's reasoning_chain (and, for conservation_reasoning, its modality_data conservation payload) genuinely and specifically supports the given answer.
quality_score (0.0-1.0): overall benchmark-item quality -- clarity, non-trivial multi-hop evidence, whether it's a good test of genuine reasoning -- independent of validity.
verification_comment: a short (5-10 word) explanation of the judgment.

Overall: 2339/2349 valid (99.6%) -- up from 72.1% (first pass) via 78.7% (conservation_reasoning methodology fix), 98.4% (interaction_propagation/evidence_attribution/disease_reasoning chain-grounding fixes), and 99.7% (v3, ClinVar placeholder-label fix) to this version (v4, disease-association confidence gate).

Validity by task_type

task_type	valid
`coding_variant`	22/22 (100.0%)
`conservation_reasoning`	160/160 (100.0%)
`counterfactual`	159/160 (99.4%)
`disease_reasoning`	244/249 (98.0%)
`evidence_attribution`	160/160 (100.0%)
`hallucination_detection`	316/320 (98.8%)
`interaction_propagation`	320/320 (100.0%)
`mechanistic_explanation`	320/320 (100.0%)
`path_traversal`	318/318 (100.0%)
`regulatory_reasoning`	160/160 (100.0%)
`structural_effect`	160/160 (100.0%)

What changed since the previous verified pass (v3)

Low-confidence GENE_ASSOCIATED_WITH_DISEASE edges (< 0.5, mostly single-source OpenTargets/KEGG associations) are no longer eligible as the answer/target in counterfactual or disease_reasoning (coding_variant Tier D) items. GENE_PARTICIPATES_IN_PATHWAY targets (curated Reactome/KEGG membership, confidence ~1.0) are unaffected. Multiple verification batches explicitly confirmed every remaining disease-answer item in the affected task types has confidence >= 0.5.

Remaining known issues (not yet fixed)

A handful of disease_reasoning items still have a disease label that duplicates the variant's own name (e.g. "BCHE Johannesburg", "APOE4 VARIANT") rather than a real condition -- a safe detection rule (disease_label == variant_label) exists but is not yet implemented; a blanket VARIANT/ISOFORM regex was checked and rejected (false positives on legitimate disease names like "behavioral variant of frontotemporal dementia").
A few hallucination_detection items have a fabricated MCQ choice expressed as a plain gene-symbol pair indistinguishable in format from the real choices, with no chain evidence singling it out specifically.
interaction_propagation's MCQ distractor generation can occasionally produce duplicate choice values for small counts (e.g. two choices both "0").
A couple of structural_effect/path_traversal items have a gene-attribution mismatch or malformed self-loop hop in the underlying KG data.
Some chain hops have a cosmetic PROTEIN/GENE node-type mislabel in bridge hops (doesn't affect answer correctness).
regulatory_reasoning items score lower on average quality_score (0.35) than other task types -- flagged for future review, not yet investigated in depth.

Source

Verification performed by independent LLM-based review (24 batches, no scripted API calls -- each batch judged directly), over iit-patna-cse-ai/GenBench_test, generated from GenBench's coding/noncoding knowledge graphs.

Downloads last month: 35

Total file size:

402 MB