PKU-Alignment/DeceptionBench
收藏Hugging Face2025-05-27 更新2025-05-31 收录
下载链接：
https://hf-mirror.com/datasets/PKU-Alignment/DeceptionBench
下载链接
链接失效反馈官方服务：
资源简介：
DeceptionBench是一个全面评估大型语言模型欺骗行为的数据集，包含180个基于实际人机交互场景的精心设计场景，分为5大欺骗类别，如奉承性误报、假装一致、隐藏能力、策略性欺骗和压力下的诚实规避。每个场景都包含内部提示、系统提示和外部提示等详细信息。
DeceptionBench is a comprehensive benchmark designed to evaluate deceptive behaviors in large language models, containing 180 meticulously crafted scenarios based on actual human-AI interaction scenarios, divided into 5 major categories of deception such as sycophantic misrepresentation, alignment faking, sandbagging, strategic deception, and honesty evasion under pressure. Each scenario includes detailed information such as inner prompt, system prompt, and outer prompt.
提供机构：
PKU-Alignment
