hltcoe/coveragebench
收藏资源简介:
CoverageBench是一个用于评估临时检索系统中信息覆盖率的统一基准套件。它关注的是“这组文档是否包含所有重要信息?”而不是“这个文档相关吗?”,衡量检索集覆盖满足信息需求所需的离散信息单元(nuggets)的完整程度。这对于检索增强生成(RAG)尤其重要,因为检索覆盖率的差距会直接导致生成答案不完整或有偏见。数据集包含7个数据集中的334个主题,每个数据集提供主题、nuggets/子主题、相关性标签和基线。具体数据集包括CAsT 2020、Fair Ranking 2022、NeuclIR 2024 Pilot、RAG 2024、RAGTIME 2025、CRUX-MultiNews和CRUX-DUC04。
CoverageBench is a unified benchmark suite for evaluating information coverage in ad hoc retrieval systems. Rather than asking is this document relevant?, coverage evaluation asks does this set of documents contain all the important information? measuring how completely a retrieved set covers the discrete information units (nuggets) needed to satisfy an information need. This is particularly important for Retrieval-Augmented Generation (RAG), where gaps in retrieval coverage propagate directly into incomplete or biased generated answers. The benchmark contains 334 topics across 7 datasets, each providing topics, nuggets/subtopics, relevance labels, and baselines. Specific datasets include CAsT 2020, Fair Ranking 2022, NeuclIR 2024 Pilot, RAG 2024, RAGTIME 2025, CRUX-MultiNews, and CRUX-DUC04.




