XRAG
收藏资源简介:
XRAG数据集是一个新型的基准,用于评估大型语言模型在跨语言检索增强生成(RAG)场景下的生成能力。该数据集由最近的新闻文章构建,确保其问题需要外部知识才能回答,并覆盖了单语和多语检索的真实世界场景。数据集包含四种语言(阿拉伯语、中文、德语和西班牙语),以及两种检索场景(单语检索和多语检索)。每个实例包含一个问题、一个答案、两篇支持文章和六篇无关文章。XRAG数据集的构建过程包括寻找相关文章对、生成跨文档问答对、质量控制、人工翻译和收集无关文章。
The XRAG dataset is a novel benchmark for evaluating the generation capabilities of large language models (LLMs) in cross-lingual retrieval-augmented generation (RAG) scenarios. This dataset is constructed from recent news articles, ensuring that its questions require external knowledge to answer, and covers real-world scenarios of monolingual and multilingual retrieval. The dataset includes four languages (Arabic, Chinese, German, and Spanish) and two retrieval scenarios: monolingual retrieval and multilingual retrieval. Each instance consists of one question, one answer, two supporting articles, and six irrelevant articles. The construction process of the XRAG dataset includes identifying relevant article pairs, generating cross-document question-answer pairs, quality control, manual translation, and collecting irrelevant articles.

- 1XRAG: Cross-lingual Retrieval-Augmented GenerationHeidelberg Institute for Theoretical Studies gGmbH, Amazon AGI · 2025年



