BioSciReview-Bench: Silent Scientific Bug Detection in Bioinformatics Code

BioSciReview-Bench evaluates coding agents on detecting, locating and classifying scientific methodology errors in bioinformatics code. The dataset contains 200 synthetic snippets in 100 correct–buggy pairs, evenly divided between Python and R. It covers five bioinformatics domains and six error categories. The release includes annotations, evaluation inputs, 1,200 predictions from six agent configurations, scores and analysis tables. This dataset accompanies BioSciReview-Bench: Benchmarking Coding Agents on Silent Scientific Bug Detection in Bioinformatics Code by Junhao Qiu, published in REALM 2026, pp. 947–956. Paper DOI: 10.18653/2026.realm-1.65. Evaluation code and usage instructions are available on GitHub. Code samples use MIT; annotations, documentation and results use CC BY 4.0. See LICENSING.md in the archive for scope.