ReleaseBench-RQA: Public Research Artifact
收藏资源简介:
ReleaseBench-RQA is the public research artifact accompanying the manuscript "From Offline Accuracy to Release Reliability: ReleaseBench-RQA, a Candidate-Level Benchmark of Multi-Metric Gates for Structured Retrieval QA." The artifact contains the data, code, cached predictions, locally sealed release decisions, model files, analysis outputs, robustness analyses, and reproducibility records used to evaluate candidate-level software release policies for structured, non-generative retrieval question answering. The primary study comprises 144 route-specific candidates. Additional analyses cover 72 joint-deployment bundles, a 48-candidate dense-retrieval architectural extension, matched-coverage and risk-coverage comparisons, contract variants, exact cluster tests, and relation-evidence-disjoint sensitivities. The study is developmental and locally decision-sealed; it was not publicly preregistered or independently replicated. No human annotation, LLM judge, human-subject data, private data, proprietary data, customer data, or employer data are included. Original documentation and derived benchmark data are licensed under CC BY 4.0. Original project code is licensed under Apache License 2.0. Third-party components retain their respective upstream licenses.



