Trazemag/hallbench
收藏Hugging Face2026-04-28 更新2026-05-03 收录
下载链接：
https://hf-mirror.com/datasets/Trazemag/hallbench
下载链接
链接失效反馈官方服务：
资源简介：
HallBench是一个用于大型语言模型幻觉检测研究的标记基准数据集。包含20,000个事实性提示，针对GPT-2（124M参数）进行测试，并标注了幻觉类型、内部激活信号和预测结果。数据集由Nikhil Upadhyay策划，包含7个知识类别，每个类别都标注了在运行GPT-2时观察到的幻觉类型，以及包括峰值事实层和抑制层在内的内部激活信号。语言为英语，许可证为MIT。
HallBench is a labeled benchmark dataset for hallucination detection research in large language models. It contains 20,000 factual prompts tested on GPT-2 (124M parameters), annotated with hallucination type, internal activation signals, and prediction outcomes. The dataset is curated by Nikhil Upadhyay and spans 7 knowledge categories, each annotated with the hallucination type observed when running GPT-2, along with internal activation signals including peak factual layer and suppression layer. The language is English, and the license is MIT.
提供机构：
Trazemag
