hsila/chem-nq

chem-nq数据集是Google Natural Questions的化学过滤子集，专为MTEB/BEIR风格的检索基准测试设计。数据源自rongzhangibm/NaturalQuestionsV2，经过过滤处理，包括基于Wikipedia化学文章标题匹配和LLM主题分类。语料库以每个保留的Wikipedia H2部分为一行，其中lead sections标记为section: "summary"，而H3-H6标题被合并到父H2文本块中，不作为text中的标题嵌入。同时，仅包含表格、画廊、参考文献、进一步阅读、外部链接和相关部分的章节被丢弃。数据集包含三个配置：corpus配置有7,874行数据，列包括_id、title、section、text；queries配置有2,745行数据，列包括_id、text；default配置有2,745行数据，列包括query-id、corpus-id、score。section列作为语料库元数据保留，以保持检索文本的清洁，同时支持分析哪个Wikipedia部分回答了查询。The chem-nq dataset is a chemistry-filtered subset of Google Natural Questions for MTEB/BEIR-style retrieval benchmarking. The data is sourced from rongzhangibm/NaturalQuestionsV2 and filtered through Wikipedia chemistry article-title matching and LLM topic classification. The corpus granularity is one row per retained Wikipedia H2 section, with lead sections using section: "summary". H3-H6 headings are merged into the parent H2 text block and are not embedded as headings in text. Table-only, gallery-only, references, further reading, external links, and see-also sections are dropped. The dataset includes three configs: corpus with 7,874 rows and columns _id, title, section, text; queries with 2,745 rows and columns _id, text; default with 2,745 rows and columns query-id, corpus-id, score. The section column is intentionally kept as corpus metadata to preserve clean retrieval text while allowing analysis of which Wikipedia section answered a query.