SearchTome
收藏资源简介:
SearchTome是由IBM创建的多领域基准数据集,旨在评估基于目录(Table of Contents)的检索能力。该数据集包含18本来自教育、金融、法律、医学、自然科学和社会科学六个领域的教科书,共计190,209条查询-目录叶节点对应样本,并已提取清晰的目录结构。创建过程利用PyMuPDF解析PDF提取目录,并以Mixtral 8x7b模型为每个段落生成多个人工问题,按比例划分为训练、开发、测试集。该数据集旨在解决长文本检索中缺乏全局语义结构的问题,促进基于目录的检索系统研究,从而提升检索精度并降低生成模型中的幻觉风险。
SearchTome is a multi-domain benchmark dataset developed by IBM, which aims to evaluate table-of-contents-based retrieval capabilities. This dataset includes 18 textbooks from six fields: education, finance, law, medicine, natural sciences, and social sciences, with a total of 190,209 query-table-of-contents leaf node corresponding samples, and clear table-of-contents structures have been extracted. During the dataset creation process, PyMuPDF was used to parse PDF files and extract table-of-contents structures, while the Mixtral 8x7b model was utilized to generate multiple artificial questions for each paragraph. The samples were then split into training, development, and test sets at a predefined ratio. This dataset is designed to address the problem of missing global semantic structure in long-text retrieval, promote research on table-of-contents-based retrieval systems, and ultimately improve retrieval accuracy while reducing hallucination risks in generative models.

- 1STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentationIBM · 2026年



