iit-patna-cse-ai/GenBench_train

GenBench CoCG QA Train Set

Train split of GenBench's multi-hop genetic reasoning QA dataset, generated from genbench-iitp/genbench-coding-qa and genbench-iitp/genbench-noncoding-qa's companion knowledge graphs, built for CoCG (Co-Evolving Confidence Graph) agent training/evaluation.

9396 items total across two configs, one per pipeline:

coding config (5693 items, pipeline coding_variant)

task_type	count
coding_variant	87
conservation_reasoning	640
counterfactual	640
disease_reasoning	493
evidence_attribution	640
hallucination_detection	640
interaction_propagation	640
mechanistic_explanation	640
path_traversal	633
structural_effect	640

noncoding config (3703 items, pipeline noncoding_regulatory)

task_type	count
disease_reasoning	503
hallucination_detection	640
interaction_propagation	640
mechanistic_explanation	640
path_traversal	640
regulatory_reasoning	640

Schema

Each item has:

• id, task_type, pipeline (coding_variant/noncoding_regulatory), difficulty
• question, answer, choices (MCQ options, when applicable)
• context -- templated chain narration
• 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, post_translational, signaling_role, 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 graphs

coding_graph.json / noncoding_graph.json are the exact knowledge graphs these items' reasoning_chain node IDs refer to -- load with GenBench's GraphBuilder.load() to resolve full node/edge attributes beyond what's inlined in each item.

Source

Generated from the coding/noncoding graphs 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. Split via scripts/split_qa_dataset.py, stratified by (pipeline, task_type), 20% test / 80% train, seed=42.

GenBench CoCG QA Train Set

Train split of GenBench's multi-hop genetic reasoning QA dataset, generated from genbench-iitp/genbench-coding-qa and genbench-iitp/genbench-noncoding-qa's companion knowledge graphs, built for CoCG (Co-Evolving Confidence Graph) agent training/evaluation.

9396 items total across two configs, one per pipeline:

coding config (5726 items, pipeline coding_variant)

task_type	count
coding_variant	105
conservation_reasoning	676
counterfactual	610
disease_reasoning	540
evidence_attribution	640
hallucination_detection	637
interaction_propagation	655
mechanistic_explanation	665
path_traversal	490
structural_effect	708

noncoding config (3670 items, pipeline noncoding_regulatory)

task_type	count
disease_reasoning	553
hallucination_detection	685
interaction_propagation	649
mechanistic_explanation	619
path_traversal	478
regulatory_reasoning	686

Schema

Each item has:

• id, task_type, pipeline (coding_variant/noncoding_regulatory), difficulty
• question, answer, choices (MCQ options, when applicable)
• context -- templated chain narration
• 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, post_translational, signaling_role, 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 graphs

coding_graph.json / noncoding_graph.json are the exact knowledge graphs these items' reasoning_chain node IDs refer to -- load with GenBench's GraphBuilder.load() to resolve full node/edge attributes beyond what's inlined in each item.

Source

Generated from the coding/noncoding graphs 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. Split via scripts/split_qa_dataset.py, stratified by (pipeline, task_type), 20% test / 80% train, seed=42 -- plus a gene-level generalization guarantee (v5): 165 genes are held out entirely from train, and 31.1% of the test set (731/2349 items) touches NO gene that appears anywhere in train -- a genuine unseen-gene slice, not just an unseen-item one. Same item content as the v4 split; only the train/test assignment changed to add this guarantee (same overall 20% test size).