Kajaluxan/version-aware-RAG-legal

Version-Aware RAG — UK Legal Corpus是一个用于检索增强生成（RAG）的版本感知评估数据集，专注于英国主要立法。该数据集包含多个时间点快照，每个法案的每个部分都以不同历史日期的版本存储，以便评估检索系统在回答版本特定和随时间变化查询方面的能力。数据集包括原始语料库（355个Markdown文件，每个文件代表一个法案部分在特定日期生效的文本）、预分块语料库（18,935个无损分块，使用与软件文档数据集相同的分块器生成）和评估集（547个经过验证的问答对，涵盖六种问题类型，如版本特定内容检索、版本列表查询、内容检索、变更检索等）。数据集基于5个英国公共一般法案（如《数据保护法2018》、《平等法2010》等），涵盖96个部分，每个部分至少有3个不同时间点版本，最多4个版本，以确保变更检索评估的有效性。数据集使用Open Government Licence v3.0许可，语言为英语（英国法律英语），总原始语料库大小约为16 MB。Version-Aware RAG — UK Legal Corpus is a version-aware evaluation dataset for Retrieval-Augmented Generation (RAG) over UK primary legislation. Every Part of every Act in the corpus is stored as multiple point-in-time snapshots — the same provision as it was in force on a given historical date — so retrieval systems can be evaluated on their ability to answer version-specific and change-over-time queries. The dataset includes a raw corpus (355 markdown files, each representing a Part of an Act as it was in force on a specific date), a pre-chunked corpus (18,935 lossless chunks generated using the same chunker as the software documentation dataset), and an evaluation set (547 validated Q&A pairs distributed across six question types, such as version-specific content retrieval, version listing & inquiry, content retrieval, change retrieval, etc.). It is based on 5 UK Public General Acts (e.g., Data Protection Act 2018, Equality Act 2010), covering 96 Parts, with each Part having at least three distinct point-in-time versions (capped at 4) to make change-retrieval evaluation meaningful. The dataset is licensed under the Open Government Licence v3.0, in English (UK legal language), with a total raw corpus size of ~16 MB.