CraneAILabs/uccb-health
收藏资源简介:
UCCB-Health Benchmark是一个基于乌干达临床指南(UCG)2023的乌干达特定临床指南基准数据集。它采用双轨结构:Track A(决策支持)主要评估乌干达临床决策支持和分诊能力,包括危险信号、转诊/升级、鉴别推理、诊断/评估、检查、特殊群体(无药物建议)以及拒绝/边界处理等类别,重点关注边界安全性;Track B(协议召回)次要评估精确的UCG治疗和剂量知识,包括一线药物、途径、频率、持续时间、剂量计划等。数据集包含1,008个项目,分布在5个文件中,涵盖自由形式问答、多项选择题和真/假陈述。数据集旨在评估临床助理模型的安全性和性能,避免不安全的处方行为。
UCCB-Health Benchmark is a Uganda-specific clinical guideline benchmark grounded in the Uganda Clinical Guidelines (UCG) 2023. It uses a multi-track structure: Track A (Decision Support) primarily measures Uganda clinical decision support and triage, covering categories such as danger signs, referral/escalation, differential reasoning, diagnosis/assessment, investigations, special populations without drug-specific advice, and refusal/boundary handling, with a focus on boundary safety; Track B (Protocol Recall) secondarily measures exact UCG treatment and dosing knowledge, including first-line drugs, routes, frequencies, durations, dosage schedules, and treatment protocol details. The dataset contains 1,008 items across 5 files, covering freeform questions, multiple-choice questions, and true/false statements. It is designed to evaluate the safety and performance of clinical assistant models, avoiding unsafe prescribing behavior.




