quotientai/HalluMix

HalluMix是一个任务无关、多领域的基准数据集，专为在现实世界的多样化场景中评估虚构检测而设计。它包含了来自不同领域（如医疗、法律、科学和新闻）和多种任务类型（如摘要、问题回答、自然语言推理）的例子。每个例子都包括文档（上下文）、回答（待评估的假设）、虚构标签（二元指示器，标记回答是否包含虚构）和来源标识（原始数据集的标签）。

HalluMix is a task-agnostic, multi-domain benchmark designed for evaluating hallucination detection in realistic, diverse contexts. It includes examples from various domains such as healthcare, law, science, and news, as well as multiple tasks like summarization, question answering, and natural language inference. Each example consists of documents (context), an answer (hypothesis to be evaluated), a hallucination label (binary indicator of whether the response contains a hallucination), and a source identifier (label for the original dataset).