OmicBench is a 44-task, biology-anchored benchmark for evaluating LLM coding agents on end-to-end multi-omics analysis workflows. Each task names a scientific objective (e.g. "identify spatial domains in DLPFC sample 151676") and a storage target by AnnData / MuData container slot -- never an API path -- and is graded end-state against biology-anchored numerical criteria.

Task structure

Tasks are organised one directory per task, in the Harbor task… See the full description on the dataset page: https://huggingface.co/datasets/omicverse/OmicBench.