David-beakr/kg-retrieval-seed
收藏资源简介:
这是一个内部使用的知识图谱检索种子数据集(版本v0.2.1),专为评估知识图谱、检索和引用能力而设计。数据集包含10个多跳问题,用于测试结构化检索(基于知识图谱)在困难的跨文档问题上是否优于扁平分块RAG,并确保答案具有可追溯的引用。每个问题都附带黄金答案、检查项(包括硬检查和奖励检查)和预期支持文档。语料库由6个固定文档组成(约670词),与配套数据集wiki-ingest-health-seed共享。数据集的设计基于STaRK和MultiHop-RAG等文献,并引入了问题类型映射和方法偏差。检查方法结合了确定性子串匹配和LLM法官评估,以分离检索质量和答案质量。数据集还支持扩展,例如增加问题数量、改进检索指标和扩大语料库规模。
This is an internal Knowledge Graph Retrieval Seed dataset (version v0.2.1) designed for evaluating knowledge graph, retrieval, and citation capabilities. It contains 10 multi-hop questions to test whether structured retrieval (based on a knowledge graph) outperforms flat chunk-based RAG on hard cross-document questions, with traceable citations. Each question includes gold answers, checks (both hard and bonus), and expected supporting documents. The corpus consists of 6 fixture documents (approximately 670 words), shared with the companion dataset wiki-ingest-health-seed. The design is grounded in literature such as STaRK and MultiHop-RAG, with question flavor mapping and methodological deviations. The check methodology combines deterministic substring matching and LLM judge evaluation to separate retrieval quality from answer quality. The dataset also supports extensions, such as adding more questions per flavor, improving retrieval metrics, and scaling to larger corpora.




