ai-safety-institute/unrelated-questions-follow-up-questions
收藏资源简介:
该数据集是一个固定的是/否后续(引发)问题集,用于无关问题谎言检测器,源自研究论文《How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions》(Pacchiardi等人,ICLR 2024)。当模型生成回应后,每个问题作为新的用户消息附加,并记录模型的是/否对数概率。每个问题的logsumexp(yes_logprobs) - logsumexp(no_logprobs)值构成特征向量,输入到训练好的探测器中。数据集包含一个CSV文件(follow_up_questions.csv),共有65个问题,其中48个在论文中使用于训练探测器。
The dataset is a fixed set of yes/no follow-up (elicitation) questions used by the Unrelated Questions lie detector, reproduced from the research paper How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions (Pacchiardi et al., ICLR 2024). After a model produces a response, each question is appended as a new user message and the models yes/no logprobs are recorded. The per-question logsumexp(yes_logprobs) - logsumexp(no_logprobs) values form the feature vector fed to the trained probe. It includes a CSV file (follow_up_questions.csv) with 65 questions in total, of which 48 are marked as used in the paper for training the probes.




