RAGBench

该数据集名为RAGBench，是首个大规模、全面的RAG基准数据集，包含了100,000个样本，跨越了五个独特的行业特定领域，并涵盖了各种RAG任务类型。RAGBench数据集还包含了可解释的标签，这些标签有助于对RAG系统进行全面的评估，并提供了可操作性的反馈，以促进系统的持续改进。其规模达到了100,000个样本，任务是对检索增强生成系统进行评估。

This dataset, named RAGBench, is the first large-scale and comprehensive RAG benchmark dataset. It includes 100,000 samples across five distinct industry-specific domains and covers a wide range of RAG task types. The RAGBench dataset also provides interpretable labels, which enable thorough evaluation of RAG systems and offer actionable feedback to support continuous system optimization. With 100,000 samples in total, this benchmark is dedicated to the evaluation of retrieval-augmented generation systems.

搜集汇总

背景与挑战

背景概述

RAGBench是一个大规模的RAG基准数据集，包含约10万示例，覆盖五个独特的行业特定领域和多种RAG任务类型，数据来源于用户手册等实际行业语料库，具有高度的行业应用相关性。数据集由12个子集组成，如covidqa和cuad，每个子集分为训练、验证和测试分割，总行数95,381行，总文件大小451 MB，采用Parquet格式和CC-BY-4.0许可证。