FACTCHD

FACTCHD是一个专为检测大型语言模型(LLMs)中事实冲突幻觉而设计的基准数据集。该数据集涵盖了多种事实性模式，包括基础、多跳、比较和集合操作，并整合了基于事实的证据链，显著增强了评估检测器解释的深度。数据集通过利用现有的知识图谱(KGs)和文本知识，采用模拟幻觉实例的方法构建，经过人工验证，确保了数据集的高效开发。FACTCHD适用于多领域评估，旨在通过其多样化的事实模式和可解释的证据链，为检测LLMs中的事实冲突幻觉设定新标准。

FACTCHD is a benchmark dataset specifically designed for detecting factual conflict hallucinations in large language models (LLMs). It covers a variety of factual patterns, including basic, multi-hop, comparative, and set operations, and integrates fact-based evidence chains, which significantly enhances the depth of evaluating the explanations provided by detection models. The dataset is constructed by leveraging existing knowledge graphs (KGs) and textual knowledge through simulating hallucinatory instances, and has undergone manual validation to ensure efficient curation of the dataset. FACTCHD is applicable to multi-domain evaluation, and aims to set a new standard for detecting factual conflict hallucinations in LLMs via its diverse factual patterns and interpretable evidence chains.