llm-semantic-router/halueval-spans-normalized

HaluEval Span-Level数据集（RAGTruth标准化提示）是一个用于幻觉检测的数据集，特别针对问答、摘要和对话任务中的幻觉跨度检测。该数据集的主要特点是将提示标准化为RAGTruth格式，以提高跨数据集的兼容性和模型训练的泛化能力。数据集包含38,711个样本，涵盖了问答、摘要和对话三种任务类型，且幻觉跨度标注平衡。每个样本包含标准化提示、回答、跨度标签、任务类型和原始提示等字段。数据集适用于训练幻觉检测器，减少数据集间的分布偏移，支持跨数据集评估研究。

The HaluEval Span-Level Dataset (RAGTruth-Normalized Prompts) is a hallucination detection dataset specifically designed for span-level hallucination detection in QA, summarization, and dialogue tasks. The dataset features prompts normalized to the RAGTruth format to improve cross-dataset compatibility and model training generalization. It contains 38,711 samples covering QA, summarization, and dialogue tasks, with balanced hallucination span annotations. Each sample includes normalized prompts, answers, span labels, task types, and original prompts. The dataset is suitable for training hallucination detectors, reducing distribution shift between datasets, and supporting cross-dataset evaluation research.