launch/FactBench
收藏Hugging Face2025-06-09 更新2025-04-08 收录
下载链接：
https://hf-mirror.com/datasets/launch/FactBench
下载链接
链接失效反馈官方服务：
资源简介：
FactBench是一个包含985个提示的213个细粒度主题的数据集，旨在捕捉语言模型在现实世界交互中的新兴事实性挑战。该数据集通过VERIFY管道识别出导致最高比例的错误或不可验证LM响应的虚构提示，并定期更新。
FactBench is a dataset of 985 prompts across 213 fine-grained topics, designed to capture emerging factuality challenges in real-world LM interactions. The dataset is composed of hallucination prompts that elicit the highest rates of incorrect or unverifiable LM responses and is regularly updated.
提供机构：
launch
