RAGTruth

该数据集是一个基准数据集，它对大型语言模型生成的回应中出现的幻觉进行了标注，涵盖了三种场景：问答（QA）、数据到文本生成（Data2Text）以及新闻摘要（Summarization）。幻觉部分在字符级别进行了标记，并且没有官方的验证集划分；随机抽取了450个实例用于验证。该数据集规模达到18,000个标注回应，任务是对 token 级别的幻觉进行检测。

This is a benchmark dataset for annotating hallucinations in responses generated by large language models, covering three scenarios: Question Answering (QA), Data-to-Text Generation (Data2Text), and News Summarization. Hallucinations are marked at the character level, and no official validation split is provided; 450 instances are randomly sampled for validation purposes. The dataset contains 18,000 annotated responses, with the task focusing on token-level hallucination detection.