SR-QA: A Test Collection for Evidence Finding in Full-Text Studies from Systematic Reviews

SR-QA (Systematic Review Question Answering) is a test collection for finding evidence in the full text of the studies included in systematic reviews. It is built from the per-study data-extraction tables of 3,037 CC BY systematic reviews in PubMed Central published between January 2025 and June 2026: each column becomes a question, the review authors' cell, copied verbatim, is the gold answer, and the CC BY full text of the cited study is where the answer should be found. No model writes a gold answer or a relevance label: a language model only describes each table's structure and phrases each column as a question, code copies the answers, and passage-level relevance labels come from string matching between the gold and the paper. The release holds 166,308 question-answer pairs over 15,944 full texts, 48,219 of which carry graded passage-level relevance labels, in five splits by each review's online date (train-2025, train-2026, dev-2026, test-2026, test-2025). It includes the full texts, table structure maps, passages, qrels and baseline runs in TREC format, a datasheet, and the code to rebuild, check and evaluate the collection. Start with README.md.