SALT-NLP/search_privacy_risk

该数据集涉及通过模拟大型语言模型（LLM）代理之间的交互来搜索隐私风险。它包含了模拟对话的详细轨迹，这些对话用于改进攻击者和防御者的指令，并探索隐私关键信息交互中的策略。数据集分为两部分：基本文件夹包含不同模型配置的基本模拟轨迹；搜索文件夹包含用于发现攻击和防御策略的搜索轨迹。

This dataset is related to searching for privacy risks in large language model (LLM) agents through simulation of dialogues. It includes detailed trajectories of simulated conversations that are used to improve the instructions of attackers and defenders, and to explore strategies in privacy-critical information exchanges. The dataset is divided into two parts: the basic folder contains basic simulation trajectories with different model configurations; the search folder contains search trajectories for discovering attack and defense strategies.