RefQA: a million-scale cross-document clinical citation-evidence Q&A dataset

RefQA is a dataset of 1,524,426 cross-document citation–evidence question–answer records extracted from clinical research articles in PubMed Central (PMC). Each record links one in-text citation from a citing paper to a cited paper (both restricted to clinical research by iCite's is_clinical flag) and pairs it with a strict-JSON CitationQA structure generated by GLM-5.1-FP8: an attributed claim, citation purpose, evidence type, support strength, cited study design, and a question–answer pair. When the cited paper supports the claim, the record includes a verbatim quote from the cited PubMed abstract, so faithfulness can be re-checked by a substring match. Structural citation extraction is deterministic from PMC JATS XML (never by a language model), confining the hallucination surface to the semantic layer. The corpus is released in two configurations: a recommended permissive subset (561,795 records; CC-BY family + public domain) and the full corpus (1,524,426 records) with a per-record license_bucket column for user-side filtering. Validation: schema validity 99.99994%; verbatim grounding faithful in 96.03% (loose) / 89.12% (exact) of quote-required records. A frozen 250-record overlap of the 1,000-record gold subset was independently annotated by two human reviewers (a PhD candidate and a board-certified faculty physician) and an LLM judge; inter-annotator agreement on the core relevance call is Cohen's κ = 0.68 and on attributed-claim faithfulness ICC = 0.72. Files in this deposit (four files; because Zenodo stores files in a flat layout, the parts that contain subfolders are packed into archives that restore their directory structure on extraction):- full_qa_clinical_v03_unique.jsonl: the full canonical corpus (1,524,426 records), one JSON object per line.- v03_unique_split_index.tsv: the deterministic train/validation/test/gold/unsplit split index, keyed by record id.- refqa_parquet_v1.0.tar.gz: the same corpus as Parquet shards; extraction yields parquet/permissive/ (the 561,795-record permissive subset) and parquet/full/ (the full corpus).- refqa_aux_v1.0.tar.gz: the validation and provenance bundle; extraction yields gold/ (gold subsets: unannotated 1,000, annotated 250 with consolidated ground truth, an adversarial confirmed-mismatch set of 10, and the frozen 250-record overlap definition), prompts/ (the frozen system prompt), and runtime/ (Technical Validation precomputed outputs, including per-record outputs from two closed-weight peer LLMs and the inter-model and inter-annotator agreement reports). License: the dataset metadata layer is released under CC-BY 4.0; source-paragraph redistribution follows the per-record license_raw / license_bucket inherited from each citing PMC article. Users requiring unambiguous commercial reuse should use the permissive subset. This deposit is the DOI-bearing canonical home of the dataset described in the companion Scientific Data Data Descriptor (DOI: to be added upon publication). Code repository: https://github.com/jin-0311/refqa. Corresponding authors: Sanghyun Park ([e-mail]), Tae Joon Jun ([e-mail]).