MinerU-Popo: Universal Post-Processing Model for Structured Document Parsing

If you like our project, please give us a star ⭐ on GitHub for the latest update.

📖 English  |  简体中文

✨ Introduction

MinerU-Popo is a lightweight and universal framework for POst-Processing OCR outputs, bridging the gap between page-level OCR parsing and document-level semantic structure. It constructs document tree structures with a 4B post-processing model that performs four subtasks: table truncation analysis, text truncation analysis, title hierarchy analysis, and image-text association analysis. We handle the challenges of cross-page geometric discontinuity, redundant document parsing, and scalability to long documents via:

• Task-Oriented Data Engine: Generate representative training data and simplify the task-specific input.
• Dynamic Chunking and Synchronization: Process long document by dynamic chunks and reduce deviations across chunks to preserve global consistency.
• Document Enrichment: Structurally construct a tree, semantically generate summaries and split long-section nodes.

📊 Performance

Better Hierarchy (TEDS) after Post-Processing

Basic OCR	Before	After
MinerU	53.7	90.6
MonkeyOCR	48.9	87.4
Dolphin	60.4	83.5
PaddleOCR	59.3	82.6
GLM-OCR	53.5	81.8

Advantages Compared to Directly Using Pre-trained Model

Model	TEDS	Doc/s
MinerU-Popo	90.6	0.37
Qwen3-VL-2B	21.2	0.22
Qwen3-VL-4B	56.5	0.20
Qwen3-VL-8B	65.9	0.16
Qwen3-VL-32B	78.0	0.04

Benefits for Downstream Retrieval and Analysis (Acc on ViDoRe V3)

Method	C.S.	Fin.	H.R.	Ind.	Phar.
MinerU-Popo	84.4	49.5	66.8	58.7	71.6
Raw RAG	82.3	48.7	63.2	60.4	64.4
Visual RAG	80.7	58.4	64.8	59.7	67.6

⚙️ Setup

1. Prepare Environment

conda create -n popo python=3.10
conda activate popo
pip install -r requirements.txt

2. Download Model

Download the MinerU-Popo post-processing model:

hf download DreamEternal/MinerU-Popo --local-dir models/Mineru-Popo

• MinerU-Popo

1. Model Configuration

In the Configuration, for transformer inference, edit the environment POPO_MODEL_PATH. For vllm inference, edit the url and key in function popo_generate.

For enrichment and question answering, further edit the url and key in qwen_generate and gpt_generate.

💻 Usage

The post-processing pipeline takes page-level parsing results from OCR/layout systems, normalizes them into a unified schema, runs MinerU-Popo inference, and finally builds document trees.

Step 1: Prepare OCR/Layout Outputs

Run your preferred page-level parser first, such as MinerU, MonkeyOCR, Dolphin, PaddleOCR-VL, or GLM-OCR. Place each model's output under:

post-process/<model_name>/

For example:

post-process/mineru/
post-process/monkeyocr/
post-process/PaddleOCR-VL-1.5/
post-process/dolphin/
post-process/glm-ocr/

Step 2: Normalize Labels

Convert raw model-specific labels and bounding boxes into the unified MinerU-Popo input format:

bash scripts/run_label_normalization.sh

The normalized outputs are written to:

outputs/label_normalization/<model_name>/

Step 3: Run MinerU-Popo Inference

Run MinerU-Popo on the normalized labels:

bash scripts/run_inference.sh

The inference outputs are written to:

outputs/inference/<model_name>/

Step 4: Build Document Trees

Build structured document trees from the inference outputs:

bash scripts/build_tree.sh

The final tree outputs and text previews are written to:

outputs/build_tree/<model_name>/
outputs/build_tree_txt/<model_name>/

Example tree outputs are provided in:

output_cases/

🙏 Acknowledgements

• MinerU and other OCR system (MonkeyOCR, Dolphin, PaddleOCR, GLM-OCR) for page-level parsing.
• ViDoRe V3 and MMDA as benchmarks.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.