atroposhealth/precision-evidence-bench

Precision Evidence Bench

Precision Evidence Bench is a Precision Medicine Benchmark from Atropos Health, the world's largest creator of real-world evidence (RWE) for clinical decision support. It evaluates how well large language models (LLMs) answer clinical questions that are grounded in patient context and inclusive of patient history: not "which treatment is better in general", but "which treatment is better for this patient", with a specific comorbidity, age, prior therapy or other characteristic that changes the answer.

A system under test is asked to do one thing: given a question a physician might actually type about a specific patient, find the evidence that answers it for that patient and report that evidence faithfully, including when no such evidence exists.

We are releasing it openly so that anyone building or evaluating clinical question-answering systems, whether grounded in the literature, in real-world data, or in web search, can test against the same questions and grade with the same rubric. Results from different systems are comparable when they use both.

What is in the benchmark

209 questions, each written against a specific comparison of two or more treatments, a patient population defined by clinical context and history, an outcome, and a follow-up window, then rendered in a clinical voice. Many questions name the patient characteristic that matters, such as a comorbidity, an age group, a lab threshold or a prior treatment, and a correct answer must match that patient rather than the general population. Some questions carry a follow-up-window instruction (0–3 or 3–12 months). Some target comparisons for which no real-world study exists, so the benchmark also measures how a system behaves when the evidence base is empty. All questions are synthetic and contain no patient data.

A five-item grading rubric (rubric.json) that scores an answer against the question and the answer's own citations, so no reference answer is needed to grade:

Rubric item	Asks whether
faithful_to_sources	the answer does not misrepresent what its cited sources show, judged as a clinician deciding whether to act on it, not as word-for-word traceability
single_source_full_pico_match	one cited source matches the question's full population, intervention, comparator and outcome, including any required patient subgroup
single_source_pic_match	one cited source matches population, intervention and comparator
single_source_pio_match	one cited source matches population, intervention and outcome
single_source_pco_match	one cited source matches population, comparator and outcome

Each item carries the exact prompt text we give to an LLM juror. Internally we use a panel of three LLM jurors from different providers and take the fraction that agree on each item; any grader that answers the five items works.

How to use it

1. Pose each question to your system exactly as written, and have it return an answer with citations.
2. Grade each answer on the five rubric items, using rubric.json as the grader's instructions.
3. Report results per item, or collapse them into a single rating. Our convention, with an item counting as true when a majority of jurors agree: full PICO match and faithful is the top rating; faithful with two, one or no partial matches step down from there; not faithful is the bottom rating regardless of matching.

Files

File	Contents
questions.csv	question_id (prwe-001 … prwe-209) and question
rubric.json	The five rubric items with their grader prompts
MANIFEST.json	Row count, SHA-256 per file, and the source commit

This release contains the questions and rubric only. Reference labels and our own model results are not included in this version.

Reporting results and contributing

If you evaluate a system on this benchmark, we would like to hear about it. Open a discussion on this dataset's page with the system evaluated, the grader used, and per-item scores. Corrections to individual questions are welcome the same way.

License and citation

Released under the Creative Commons Attribution-ShareAlike 4.0 International license (cc-by-sa-4.0): use, share and adapt it freely, with attribution to Atropos Health, and release derivatives under the same license.

Please cite the dataset as Atropos Health, Precision Evidence Bench: a Precision Medicine Benchmark for evidence-based clinical question answering (2026), https://huggingface.co/datasets/atroposhealth/precision-evidence-bench.

Provenance

Built by scripts/build_hf_release.py in Atropos Health's prwe_bench repository; MANIFEST.json records the source commit and file hashes. Built 2026-09-17.