SaTML LLM Capture-the-Flag Competition Dataset

SaTML LLM Capture-the-Flag Competition Dataset是由苏黎世联邦理工学院等机构创建的一个大型数据集，包含超过137,000条多轮攻击聊天记录。该数据集主要来源于2024年IEEE SaTML大型语言模型捕获标志竞赛的攻击阶段，旨在研究大型语言模型系统面临的安全风险。数据集不仅包含简单的提示和规则，还包括复杂的过滤器和多轮对话，适用于研究LLM系统的安全性。通过公开此数据集和竞赛代码库，旨在促进未来对LLM安全性的研究。

The SaTML LLM Capture-the-Flag Competition Dataset is a large-scale dataset developed by ETH Zurich and other institutions, containing over 137,000 multi-turn adversarial chat records. Mainly sourced from the adversarial stage of the 2024 IEEE SaTML Large Language Model Capture-the-Flag Competition, this dataset is designed to investigate the security risks faced by large language model (LLM) systems. In addition to simple prompts and rules, the dataset also encompasses complex filters and multi-turn dialogues, making it applicable for research on LLM system security. By publicly releasing this dataset and the competition codebase, the project aims to facilitate future research on LLM security.