← All benchmarks

micro1 PharmaBench

Health & Medicine

micro1's PharmaBench, unrelated to the ADMET PharmaBench dataset: 120 clinical-development documentation tasks written by biostatisticians, medical writers, safety physicians and regulatory specialists inside eight synthetic drug programs. Each task gives an agent the program's documents to date, often more than a hundred PDFs and datasets, in an offline sandbox with a shell and Python, and an LLM judge grades the answer against the expert's weighted rubric with a reference answer as context. A run passes when it meets every required criterion and triggers no penalty; the score is the share of passing runs over three runs per task.

Top models (higher is better)