MisBench

该数据集名为MisBench，是评估大型语言模型（LLMs）在面对错误信息时的行为及其知识偏好的最大且最全面的基准测试。它包含了1034万6708条错误信息，这些信息涵盖了知识性冲突和风格上的变异。实证结果显示，大型语言模型会受到错误信息中的知识冲突和风格变异的影响。该数据集的任务是评估LLMs在面对错误信息时的行为和知识偏好。

This dataset, named MisBench, is the largest and most comprehensive benchmark for evaluating the behaviors and knowledge preferences of Large Language Models (LLMs) when exposed to misinformation. It contains 10,346,708 pieces of misinformation covering both knowledge conflicts and stylistic variations. Empirical results demonstrate that LLMs are susceptible to the influence of knowledge conflicts and stylistic variations present in such misinformation. The core task of this benchmark is to assess the behaviors and knowledge preferences of LLMs when confronted with misinformation.