French PDF-to-Markdown Benchmark
收藏资源简介:
该数据集由Probayes和OpenValue等机构联合构建,专注于法语PDF文档到Markdown格式的转换评估。数据集包含约60,000份文档,涵盖手写表单、复杂布局、密集表格和图形丰富页面等多种挑战性内容,通过模型分歧采样策略筛选困难样本。数据主要来源于CCPDF和Gallica语料库,经过半自动化测试生成和人工验证流程处理。该基准测试旨在解决多语言文档理解中的实际痛点,特别针对法语文档在检索增强生成(RAG)流程中的语义完整性和结构连贯性需求,为视觉语言模型在复杂现实场景下的性能评估提供标准化工具。
This dataset was jointly constructed by institutions including Probayes, OpenValue, and other relevant organizations, focusing on the evaluation of French PDF document conversion to Markdown format. The dataset contains approximately 60,000 documents covering a wide range of challenging content types such as handwritten forms, complex layouts, dense tables, and graphics-rich pages. Difficult samples were screened using a model disagreement sampling strategy. The data is primarily sourced from the CCPDF and Gallica corpora, and processed through semi-automated test generation and manual verification workflows. This benchmark aims to address the practical pain points in multilingual document understanding, specifically targeting the requirements for semantic integrity and structural coherence of French documents in Retrieval-Augmented Generation (RAG) workflows, providing a standardized tool for evaluating the performance of vision-language models in complex real-world scenarios.
vlmparse 数据集概述
数据集基本信息
- 数据集名称: vlmparse
- 数据集托管平台: Hugging Face
- 数据集访问地址: https://huggingface.co/datasets/pulsia/fr-bench-pdf2md
- 相关基准测试: Benchmark (https://github.com/ld-lab-pulsia/benchpdf2md)
- 相关排行榜: Leaderboard (https://huggingface.co/spaces/pulsia/fr-bench-pdf2md)
数据集核心功能
vlmparse 是一个统一的封装工具,用于结合视觉语言模型和OCR解决方案,将PDF文档解析为Markdown格式。
主要特性
- ⚡ 支持异步/并发处理以实现高吞吐量
- 🐳 为本地模型提供自动Docker服务器管理
- 🔄 跨所有VLM/OCR提供商的统一接口
- 📊 内置基于Streamlit的结果可视化功能
支持的转换器类型
开源小型VLM
lightonocr2, mineru2.5, hunyuanocr, paddleocrvl-1.5, granite-docling, olmocr2-fp8, dotsocr, chandra, deepseekocr2, nanonets/Nanonets-OCR2-3B
开源通用VLM
例如Qwen系列模型
处理流水线
docling
专有大型语言模型
gemini, gpt
开发团队

- 1Benchmarking Vision-Language Models for French PDF-to-Markdown ConversionProbayes; OpenValue · 2026年



