DTBench
收藏资源简介:
DTBench是由浙江大学和香港科技大学广州校区联合开发的合成基准数据集,旨在系统评估文档到表格(Doc2Table)提取任务中大型语言模型的多维能力。该数据集包含120个案例共8,811个单元格级评估实例,采用逆向Table2Doc范式生成,通过多智能体合成工作流注入五种核心能力维度(如转换对齐、推理归纳等)的差异化难度。数据来源为人工构建的真实表格反向生成的合成文档,严格覆盖了格式转换、冲突解决等13种子能力场景。该数据集主要应用于自然语言处理和数据库交叉领域,为提升结构化信息抽取的准确性、可验证性提供标准化测试平台。
DTBench is a synthetic benchmark dataset jointly developed by Zhejiang University and The Hong Kong University of Science and Technology (Guangzhou). It aims to systematically evaluate the multi-dimensional capabilities of large language models (LLMs) in the document-to-table (Doc2Table) extraction task. This dataset includes 120 cases with a total of 8,811 cell-level evaluation instances, and is generated using the reverse Table2Doc paradigm. Through a multi-agent synthetic workflow, it introduces differentiated difficulty levels across five core capability dimensions such as conversion alignment, reasoning and induction, etc. The dataset is sourced from synthetic documents generated by reversing manually constructed real tables, which strictly covers 13 sub-capability scenarios including format conversion and conflict resolution. This dataset is primarily applied in the cross-disciplinary field of natural language processing (NLP) and databases, serving as a standardized test platform to improve the accuracy and verifiability of structured information extraction.
- 1DTBench: A Synthetic Benchmark for Document-to-Table Extraction浙江大学; 香港科技大学·广州 · 2026年



