SynthPAI

SynthPAI是由苏黎世联邦理工学院创建的一个合成数据集，包含超过7800条手动标注的个人属性推断标签的评论。该数据集通过模拟流行社交媒体平台如Reddit的评论线程环境，使用个性化LLM代理生成。SynthPAI旨在解决大型语言模型在个人属性推断方面的隐私威胁，通过提供一个隐私保护的基础，支持未来关于理解和缓解基于推断的隐私威胁的研究。数据集涵盖广泛的年龄、性别、收入等个人属性，并已被验证在个人属性推断研究中具有代表性。

SynthPAI is a synthetic dataset created by ETH Zurich, containing over 7,800 manually annotated comments with labels for personal attribute inference. Generated using personalized LLM agents, this dataset simulates the comment thread environments of popular social media platforms such as Reddit. SynthPAI aims to address the privacy threats posed by large language models (LLMs) in the context of personal attribute inference, providing a privacy-preserving benchmark to support future research on understanding and mitigating inference-based privacy threats. The dataset covers a wide range of personal attributes including age, gender, income, and has been validated to be representative for personal attribute inference research.