--- title: MonkeyDoc canonical_url: "https://www.modelscope.cn/datasets/zenosai/MonkeyDoc" md_url: "https://www.modelscope.cn/datasets/zenosai/MonkeyDoc.md" repository: zenosai/MonkeyDoc chinese_name: MonkeyDoc last_updated: 2026-01-30 license: "Apache License 2.0" storage_size: "1.6 TB" downloads: 112692 stars: 10 --- # MonkeyDoc > MonkeyDoc - zenosai 在 ModelScope 开源的数据集。文档解析数据集MonkeyDoc zenosai/MonkeyDoc 是 ModelScope 魔搭社区上的数据集，存储大小 1.6 TB，采用 Apache License 2.0 许可。 - **Repository**: zenosai/MonkeyDoc - **License**: Apache License 2.0 - **Storage size**: 1.6 TB - **Downloads**: 112692 - **Stars**: 10 - **Last updated**: 2026-01-30 Source: https://www.modelscope.cn/datasets/zenosai/MonkeyDoc --- ## Download and Extract ### 1) Download the dataset Download the dataset to a local directory, for example: ````bash modelscope download --repo-type dataset zenosai/MonkeyDoc --local-dir /path/to/MonkeyDoc ```` ### 2) Extract all image archives ````bash cd /path/to/MonkeyDoc # Extract all .tar.zst archives under images into the images folder for f in ./images/*.tar.zst; do tar -I zstd -xf "$f" -C . done ```` ### 3) (Optional) Add prefix to all image paths in MonkeyDoc.json If needed, add the prefix `/path/to/MonkeyDoc/` to all image paths in `MonkeyDoc.json`. ````bash python add_prefix.py \ --input /path/to/MonkeyDoc/MonkeyDoc.json \ --output /path/to/MonkeyDoc/MonkeyDoc_prefixed.json \ --prefix /path/to/MonkeyDoc/ ````

Necessary details of the MonkeyDoc data generation pipeline

### 1. Text, Formula, and Table Recognition Data Generated by Qwen2.5VL-72B and Gemini-2.5-Pro We use Qwen2.5-VL-72B to generate text and formula data. Since its table recognition performance does not meet our requirements, we rely exclusively on Gemini-2.5 Pro for table data generation. Here, we present the prompts used for recognition. #### Text & Formula Recognition **Model:** Qwen2.5VL-72B ```python # Text Recognition Prompt text_prompt = ( "Please output the text content from the image. Inline formulas should be wrapped with $ " "and fully retain the format of the mathematical formulas. Do not wrap non-mathematical " "content with $. Avoid line breaks within sentences, but allow line breaks after complete " "sentences when appropriate for text structure (e.g., paragraph separation, list items, " "or logical section breaks)." ) # Formula Recognition Prompt formula_prompt = ( "Please write out the expression of the formula in the image using LaTeX format. " "If there is a corresponding number for the formula, generate it as well. " "Do not output content unrelated to the formula in the image." ) ``` #### Table Recognition **Model:** Gemini 2.5 Pro ```python # Table Recognition Prompt table_prompt = ( "你是HTML编码助手，只输出纯

	代码，完全反映表格图像结构，" "无