nlpai-lab/mirage
收藏Hugging Face2025-05-18 更新2025-05-31 收录
下载链接：
https://hf-mirror.com/datasets/nlpai-lab/mirage
下载链接
链接失效反馈官方服务：
资源简介：
MIRAGE是一个用于评估检索增强生成（Retrieval-Augmented Generation, RAG）系统的基准数据集，包含7560个问答对和从各种基于维基问答数据集中精选的37800个上下文池。这个数据集能够对大型语言模型和检索器在现实、噪声和理想化环境下的表现进行稳健评估，并引入了新的指标来分析上下文敏感性、噪声脆弱性和检索有效性。
MIRAGE is a benchmark dataset for evaluating Retrieval-Augmented Generation (RAG) systems, featuring 7,560 QA pairs and 37,800 context pools curated from diverse Wikipedia-based QA datasets (IfQA, NaturalQA, TriviaQA, DROP, PopQA). MIRAGE enables robust assessment of LLMs and retrievers under realistic, noisy, and oracle settings, and introduces novel metrics for analyzing context sensitivity, noise vulnerability, and retrieval effectiveness.
提供机构：
nlpai-lab
