RAG vs. No-RAG benchmark dataset: LLM-judged clinical reasoning in transplant infectious disease

Dataset and code supporting “A Hybrid Retrieval-Augmented Generation Configuration Lowered Clinical Answer Quality Across 39 Language Models” (PLOS Digital Health, PDIG-D-26-01326; originally submitted as “Retrieval-Augmented Generation Reduces Synthesis-Intensive Clinical Reasoning Quality in Transplant Infectious Disease”). Version 2 adds the revision files in revision_release_v2.zip. The version 1 files are unchanged: all_models_rag_non_rag_judged.csv: 4,911 LLM-judge records from 1,637 model outputs (39 models, 21 transplant infectious disease vignettes, and two configurations, RAG and No-RAG; one No-RAG output was missing because of a generation failure). Each output was scored by three LLM judges on six clinical-reasoning dimensions; the file also includes generation-time grounding metrics and major-error annotations. expert_evaluations_linked_long.csv: 60 answer-level ratings from the original three evaluators on 10 paired RAG/No-RAG cases (accuracy, usefulness, major errors, preference). Condition and model labels were masked, but citations and answer formatting could reveal the configuration. retrieval_quality_classification.csv: document-level classification of the 21 cases into six retrieval-quality categories, with similarity scores. revision_release_v2.zip contains: revision_data/: the 21 case prompts; 100 answer-level ratings from five evaluators (E1–E5) on the ten shared case pairs; evaluator- and case-level differences; S1–S4 Tables as CSV. revision_statistics/: output-averaged estimates clustered by model and case, dimension-level FDR results, judge-specific estimates, sensitivity analyses, and case-level expert results. revision_scripts/: analysis, figure, table, and passage-inspection code, with the run order in README.md. retrieval/: passage manifest, source catalogue, passage access guide, audit rubric, coverage table, and correction ledger. Retrieved passage text is not redistributed because the source documents are copyrighted. Code: https://github.com/KeatingLabNYU/rag-txid-benchmark