刚刚，超1亿文档图像数据开源：MonkeyOCRv2想给视觉文档补上一块公共底座
Merged summary
TL;DR - MonkeyOCRv2 introduces a document-native visual encoder trained with text generation and pixel reconstruction to preserve character- and layout-level evidence. Its team also released MonkeyDoc v2, a 100+ TB dataset containing 113 million document images across 17 languages.
- Pixel reconstruction improved shuffled-text recognition at low resolution from 55.4% to 72.1%, reducing reliance on language-model guessing.
- With a frozen Qwen3-1.7B backend, MonkeyOCRv2-B averaged 57.2 across eight document-understanding benchmarks versus 44.0 for OpenVision-B.
- A 0.7B document parser built around the frozen encoder scored 83.3 on multilingual MDPBench, 2.8 points above a prior 3B open-source baseline.
- MonkeyDoc v2 includes 8 million full pages and 105 million cropped elements, combining 61 million real-world and 52 million synthetic samples.
Sources (1)
刚刚，超1亿文档图像数据开源：MonkeyOCRv2想给视觉文档补上一块公共底座
TL;DR - MonkeyOCRv2 introduces a document-native visual encoder trained with text generation and pixel reconstruction to preserve character- and layout-level evidence. Its team also released MonkeyDoc v2, a 100+ TB dataset containing 113 million document images across 17 languages.
- Pixel reconstruction improved shuffled-text recognition at low resolution from 55.4% to 72.1%, reducing reliance on language-model guessing.
- With a frozen Qwen3-1.7B backend, MonkeyOCRv2-B averaged 57.2 across eight document-understanding benchmarks versus 44.0 for OpenVision-B.
- A 0.7B document parser built around the frozen encoder scored 83.3 on multilingual MDPBench, 2.8 points above a prior 3B open-source baseline.
- MonkeyDoc v2 includes 8 million full pages and 105 million cropped elements, combining 61 million real-world and 52 million synthetic samples.