SOSBench/SOSBench

SOSBench是一个用于学术研究的学术数据集，主要目的是用于大型语言模型的安全性、对齐、监督、红队和政策分析的研究。次要目的是用于课堂教学和演示安全攻击/防御（研究生级别或以上）。数据集包含约3000个示例，分为训练集。它涵盖了六个高风险领域：化学、生物学、医学、物理、心理学和生物安全。数据集的收集方法是通过对源自规定的种子术语进行覆盖驱动的突变和专家评审来发展提示。数据集的用途受到严格的限制，禁止用于任何可能造成伤害的活动，如设计、促进或评估现实世界中的有害活动。

SOSBench is an academic dataset for research purposes, primarily intended for research on the safety, alignment, oversight, red-teaming, and policy analysis of large language models. The secondary purpose is for classroom teaching and demonstration of safety attacks/defences at the graduate level or above. The dataset contains approximately 3000 examples, split into a training set. It covers six high-risk areas: chemistry, biology, medicine, physics, psychology, and biosecurity. The collection method of the dataset involves evolving prompts from regulation-derived seed terms using coverage-driven mutation and expert review. The use of the dataset is strictly regulated, prohibiting any activities that could cause harm, such as designing, facilitating, or evaluating harmful activities in the real world.