CodeSoulco/TextInsightBench
TextInsightBench English | 简体中文 A natural-language data-mining benchmark for agents: 50 tasks, 435,000 task documents and 944,468 unlabeled learning documents. Each task provides 5,000 or 10,000 texts and a research objective. Agents choose the patterns, populations and comparisons to investigate, then submit up to three findings with complete document assignments, exact quotations, statistics, counterexamples and limitations. Any analysis method is allowed. Contents… See the full description on the dataset page: https://huggingface.co/datasets/CodeSoulco/TextInsightBench.
0970
1{2 "version": "textinsightbench",3 "evaluation_unit": "evidence-backed downstream finding",4 "discovery_mode": "agent_selected",5 "scoring_version": "finding-quality-discovery-evidence",6 "task_families": {7 "group_difference": 20,8 "temporal_change": 15,9 "compound_association": 1510 },11 "tracks": {12 "task_only": "Use the supplied task corpus.",13 "unlabeled_pool": "Learn from the released pool before evaluation."14 },15 "reference_policy": "No fixed reference conclusions; score against corpus evidence and the public rubric.",16 "semantic_review": "Sampled claim-blind document checks cap narrative quality; exhaustive arithmetic, sampled semantic verification.",17 "independent_validation": false,18 "historical_exposure": "Task documents were drawn from a previously public learning pool. Not an unseen or independent validation set.",19 "cross_task_adaptation": "Freeze global prompts, learned parameters and thresholds; no evaluation-feedback transfer.",20 "security": "Untrusted corpus content. Process runner is not a sandbox."21}22