AutoHall

AutoHall是一个自动生成的幻觉检测数据集，用于大型语言模型。该数据集利用现有的事实核查数据集，通过自动构建模型特定的幻觉数据集来解决手动标注幻觉生成内容的耗时和成本问题。数据集内容涵盖多个领域，如科技、文化、健康等，旨在通过分析不同模型在不同主题上的幻觉生成倾向，提高幻觉检测的准确性。创建过程涉及从事实核查数据集中提取声明，生成相关参考，并通过比较参考与事实标签来确定幻觉是否发生。AutoHall的应用领域主要集中在提高大型语言模型的可靠性和用户信任，特别是在生成任务中减少不准确或虚构信息的产生。

AutoHall is an automatically generated hallucination detection dataset tailored for large language models. Leveraging existing fact-checking datasets, it constructs model-specific hallucination datasets automatically, which solves the time-consuming and cost-intensive issues of manually annotating hallucinatory generated content. Covering multiple domains such as technology, culture, health and others, this dataset is designed to analyze the hallucinatory generation tendencies of different models across various topics, thereby improving the accuracy of hallucination detection. Its development workflow includes extracting claims from fact-checking datasets, generating corresponding reference materials, and verifying the presence of hallucinations by comparing these references with the attached factual labels. The main application areas of AutoHall focus on boosting the reliability and user trust of large language models, especially reducing the generation of inaccurate or fictional information in generative tasks.