genbench-iitp/coding-variant

GenBench CoCG QA Dataset

Multi-hop genetic reasoning QA items generated from GenBench's knowledge graph (Ensembl, ClinVar, VEP, BioGRID, STRING, Reactome, UniProt, GO, SIGNOR, OmniPath, KEGG, DisGeNET, OpenTargets, PubTator3, GTEx, and more), built for CoCG (Co-Evolving Confidence Graph) agent training.

2513 items across 11 task types.

Task types

Schema

Each item has:

• —id, task_type, pipeline (coding_variant/noncoding_regulatory), difficulty
• —question, answer, choices (MCQ options, when applicable)
• —context -- either a templated chain narration, or (if llm_rewrite was applied) an LLM-rewritten fluent Step/Evidence/Interpretation/Conclusion narrative
• —reasoning_chain -- the grounded, machine-checkable multi-hop path (steps: each with source_node_id/target_node_id/edge_relation/ edge_confidence/edge_source_db), never touched by any LLM step
• —modality_data -- raw modality payloads (sequence, structural, transcriptomic, posttranslational, signalingrole, etc.) attached to the chain's anchor nodes
• —evidence -- supporting evidence entries with source database/PMID
• —path_confidence_score -- continuous, confidence-derived difficulty score

Companion graph

graph.json (if included in this repo) is the exact knowledge graph these items' reasoning_chain node IDs refer to -- load it with GenBench's `GraphBuilder.load()` to resolve full node/edge attributes beyond what's inlined in each item.

Source

Generated from data\curated\qa_dataset.jsonl in GenBench, the substrate for CoCG (Co-Evolving Confidence Graph) agent training -- per-edge, per-modality KG confidence that co-adapts with an RL policy during training rather than treating the KG as a frozen oracle.