Medical Reasoning

Purpose & provenance

Chinese synthetic rewrites from HealthBench Professional source families: 29 cases, reference answers, 536 weighted rubric items and 125 recorded evidence references. Both JSONL formats contain the same cases, with source attribution and modification notes.

Public sample

Download dataset ↗

Schema / fields & files

Field and type excerpts; consult the pinned source revision and original documentation.

id / question / answer_format

string: stable case ID, Chinese task and required response structure; question and answer_format together form the input.

standard_answer

string: reference answer; keep separate from the evaluated model input.

rubric

array[object]: criterion is string, weight is int, critical is bool, and acceptable_alternatives is string.

evidence / source

array[object] / object: evidence references and source revision, original IDs and family mapping. Recorded summaries are not a new clinical literature review.

healthbench/medical_eval.jsonl

One JSON object per line: id, conversation.messages and rubric_items; the same cases as default, not an additional independent dataset.