ai-safety-institute/AgentHarm

AgentHarm是一个用于测量大型语言模型（LLM）代理有害性的基准数据集。数据集包含三种配置：无害良性（harmless_benign）、有害（harmful）和聊天（chat），每种配置都有公开测试和验证数据文件。数据集旨在支持研究，使AI系统对所有人更安全，包含可能表达有害情感的内容。数据集的使用受到MIT许可证的限制，并附加了额外条款，禁止将数据集用于提高AI系统安全性和安全性之外的目的。

AgentHarm is a benchmark dataset for measuring the harmfulness of large language model (LLM) agents. The dataset includes three configurations: harmless_benign, harmful, and chat, each with public test and validation data files. The dataset is designed to support research that makes AI systems safer for everyone, containing content that may express harmful sentiments. The use of the dataset is governed by the MIT License with an additional clause that prohibits using the dataset and benchmark for purposes other than improving the safety and security of AI systems.