med-harm

该数据集是为了评估大型语言模型（LLM）的医疗安全性而生成的一系列有害医疗提示，其依据是美国医学协会提出的九大医学伦理原则。该数据集采用了两种公开可用的LLM——GPT-4和Llama-2，并通过不同的生成技术构建而成。数据集规模包括1,742个有害医疗提示，均匀分为训练集和测试集，旨在执行评估LLM医疗安全性的任务。

This dataset is a collection of harmful medical prompts generated to evaluate the medical safety of Large Language Models (LLMs), grounded on the nine medical ethics principles proposed by the American Medical Association. It was constructed using two publicly available LLMs, GPT-4 and Llama-2, with a variety of distinct generation techniques. Comprising 1,742 harmful medical prompts in total, the dataset is evenly split into a training set and a test set, and it is designed for the task of evaluating the medical safety of LLMs.