phylobio/BixBench-Verified-50

BixBench-Verified-50 A curated subset of BixBench with 50 verified questions across 33 unique data capsules, designed for reliable evaluation of AI agents on computational biology tasks. Overview BixBench-Verified-50 was created to isolate real agent performance from benchmark issues. Starting from the full BixBench benchmark, we sampled questions and identified problematic ones. Some were removed entirely. For others, we revised the question text for clarity or… See the full description on the dataset page: https://huggingface.co/datasets/phylobio/BixBench-Verified-50.