LongDocURL
收藏资源简介:
LongDocURL是一个由中国科学院自动化研究所和阿里巴巴淘宝天猫集团联合创建的多模态长文档基准数据集,旨在评估模型在处理复杂文档元素、长上下文和多样化任务中的能力。该数据集包含2325个高质量问答对,覆盖了超过33,000页的文档,涵盖了理解、推理和定位三大任务,并细分为20个子任务。数据集的构建采用了半自动化的流程,包括文档筛选、问答生成和自动化与人工验证等步骤。LongDocURL的应用领域主要集中在文档理解、数值推理和跨元素定位等任务,旨在解决现有基准在处理长文档和复杂元素时的不足,推动文档理解领域的发展。
LongDocURL is a multimodal long-document benchmark dataset jointly developed by the Institute of Automation, Chinese Academy of Sciences and Alibaba Taobao Tmall Group. It is designed to evaluate the performance of models when dealing with complex document elements, long contexts and diverse tasks. This dataset comprises 2,325 high-quality question-answer pairs, covering documents with a total of over 33,000 pages, and includes three core tasks: comprehension, reasoning and localization, which are further subdivided into 20 subtasks. The construction of LongDocURL follows a semi-automated pipeline, encompassing steps such as document screening, question-answer generation, as well as automated and manual verification. The primary application scenarios of LongDocURL concentrate on document understanding, numerical reasoning and cross-element localization tasks. This benchmark aims to mitigate the limitations of existing benchmarks in handling long documents and complex elements, thereby advancing the development of the document understanding domain.
LongDocURL 数据集概述
数据集简介
- 名称: LongDocURL
- 用途: 评估模型在长文档理解、推理和定位方面的能力。
- 数据量: 包含 2,325 个高质量的问答对,覆盖 396 个 PDF 格式的文档,总计超过 33,000 页。
- 数据格式: PDF 文件和问答对文件(.jsonl)。
数据集获取
- 数据集地址: LongDocURL
- 相关资源:
评估方法
-
下载与提取 PDF 文件:
- 从 LongDocURL 下载 PDF 和问答对文件。
- 使用 PyMuPDF 提取 PDF 文件为 png 和 json 文件,运行以下命令: bash bash utils/run_extract_ccpdf.sh
-
评估 API 模型:
- 运行以下命令进行评估: bash bash scripts/eval_api_models.sh
排行榜
| 模型 | 大小 | 理解能力 | 推理能力 | 定位能力 | 总分 |
|---|---|---|---|---|---|
| GPT-4o-24-05-13 🥇 | - | 68.6 | 59.9 | 59.6 | 64.5 |
| Gemini-1.5-Pro 🥈 | - | 55.7 | 43.4 | 46.4 | 50.9 |
| Qwen-VL-Max 🥉 | - | 58.8 | 43.9 | 36.0 | 49.5 |
| Qwen2-VL | 7B | 36.9 | 24.8 | 22.6 | 30.6 |
| LLaVA-OneVision-Chat | 7B | 30.5 | 19.0 | 18.7 | 25.0 |
| LLaVA-Next-Interleave-DPO | 7B | 21.6 | 13.9 | 7.6 | 16.2 |
| Llama-3.2 | 11B | 12.9 | 9.4 | 2.7 | 9.2 |




