Modelpublic

Logics-MLLM/Logics-Parsing

sourceHugging Faceupdated 1y agoView on Hugging Face

51likes255downloads

Model Card

🤗 <a href="https://github.com/alibaba/Logics-Parsing">GitHub</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://www.modelscope.cn/studios/Alibaba-DT/Logics-Parsing/summary">Demo</a>&nbsp&nbsp | &nbsp&nbsp📑 <a href="https://arxiv.org/abs/2509.19760">Technical Report</a>

Introduction

<div align="center"> <table style="width: 800px;"> <tr> <td align="center"> <img src="imgs/report.gif" alt="研报示例"> </td> <td align="center"> <img src="imgs/chemistry.gif" alt="化学分子式示例"> </td> <td align="center"> <img src="imgs/paper.gif" alt="论文示例"> </td> <td align="center"> <img src="imgs/handwritten.gif" alt="手写示例"> </td> </tr> <tr> <td align="center">report</td> <td align="center">chemistry</td> <td align="center">paper</td> <td align="center">handwritten</td> </tr> </table> </div>

Logics-Parsing is a powerful, end-to-end document parsing model built upon a general Vision-Language Model (VLM) through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). It excels at accurately analyzing and structuring highly complex documents.

Key Features

—Effortless End-to-End Processing
—Our single-model architecture eliminates the need for complex, multi-stage pipelines. Deployment and inference are straightforward, going directly from a document image to structured output.
—It demonstrates exceptional performance on documents with challenging layouts.

—Advanced Content Recognition
—It accurately recognizes and structures difficult content, including intricate scientific formulas.
—Chemical structures are intelligently identified and can be represented in the standard SMILES format.

—Rich, Structured HTML Output
—The model generates a clean HTML representation of the document, preserving its logical structure.
—Each content block (e.g., paragraph, table, figure, formula) is tagged with its category, bounding box coordinates, and OCR text.
—It automatically identifies and filters out irrelevant elements like headers and footers, focusing only on the core content.

—State-of-the-Art Performance
—Logics-Parsing achieves the best performance on our in-house benchmark, which is specifically designed to comprehensively evaluate a model’s parsing capability on complex-layout documents and STEM content.

Benchmark

Existing document-parsing benchmarks often provide limited coverage of complex layouts and STEM content. To address this, we constructed an in-house benchmark comprising 1,078 page-level images across nine major categories and over twenty sub-categories. Our model achieves the best performance on this benchmark. <div align="center"> <img src="imgs/BenchCls.png"> </div> <table> <tr> <td rowspan="2">Model Type</td> <td rowspan="2">Methods</td> <td colspan="2">Overall Edit ↓</td> <td colspan="2">Text Edit Edit ↓</td> <td colspan="2">Formula Edit ↓</td> <td colspan="2">Table TEDS ↑</td> <td colspan="2">Table Edit ↓</td> <td colspan="2">ReadOrderEdit ↓</td> <td rowspan="1">ChemistryEdit ↓</td> <td rowspan="1">HandWritingEdit ↓</td> </tr> <tr> <td>EN</td> <td>ZH</td> <td>EN</td> <td>ZH</td> <td>EN</td> <td>ZH</td> <td>EN</td> <td>ZH</td> <td>EN</td> <td>ZH</td> <td>EN</td> <td>ZH</td> <td>ALL</td> <td>ALL</td> </tr> <tr> <td rowspan="7">Pipeline Tools</td> <td>doc2x</td> <td>0.209</td> <td>0.188</td> <td>0.128</td> <td>0.194</td> <td>0.377</td> <td>0.321</td> <td>81.1</td> <td>85.3</td> <td><ins>0.148</ins></td> <td><ins>0.115</ins></td> <td>0.146</td> <td>0.122</td> <td>1.0</td> <td>0.307</td> </tr> <tr> <td>Textin</td> <td>0.153</td> <td>0.158</td> <td>0.132</td> <td>0.190</td> <td>0.185</td> <td>0.223</td> <td>76.7</td> <td><ins>86.3</ins></td> <td>0.176</td> <td>0.113</td> <td>0.118</td> <td>0.104</td> <td>1.0</td> <td>0.344</td> </tr> <tr> <td>mathpix*</td> <td><ins>0.128</ins></td> <td><ins>0.146</ins></td> <td>0.128</td> <td><ins>0.152</ins></td> <td>0.06</td> <td>0.142</td> <td>86.2</td> <td>86.6</td> <td>0.120</td> <td>0.127</td> <td>0.204</td> <td>0.164</td> <td>0.552</td> <td>0.263</td> </tr> <tr> <td>PP_StructureV3</td> <td>0.220</td> <td>0.226</td> <td>0.172</td> <td>0.29</td> <td>0.272</td> <td>0.276</td> <td>66</td> <td>71.5</td> <td>0.237</td> <td>0.193</td> <td>0.201</td> <td>0.143</td> <td>1.0</td> <td>0.382</td> </tr> <tr> <td>Mineru2</td> <td>0.212</td> <td>0.245</td> <td>0.134</td> <td>0.195</td> <td>0.280</td> <td>0.407</td> <td>67.5</td> <td>71.8</td> <td>0.228</td> <td>0.203</td> <td>0.205</td> <td>0.177</td> <td>1.0</td> <td>0.387</td> </tr> <tr> <td>Marker</td> <td>0.324</td> <td>0.409</td> <td>0.188</td> <td>0.289</td> <td>0.285</td> <td>0.383</td> <td>65.5</td> <td>50.4</td> <td>0.593</td> <td>0.702</td> <td>0.23</td> <td>0.262</td> <td>1.0</td> <td>0.50</td> </tr> <tr> <td>Pix2text</td> <td>0.447</td> <td>0.547</td> <td>0.485</td> <td>0.577</td> <td>0.312</td> <td>0.465</td> <td>64.7</td> <td>63.0</td> <td>0.566</td> <td>0.613</td> <td>0.424</td> <td>0.534</td> <td>1.0</td> <td>0.95</td> </tr> <tr> <td rowspan="8">Expert VLMs</td> <td>Dolphin</td> <td>0.208</td> <td>0.256</td> <td>0.149</td> <td>0.189</td> <td>0.334</td> <td>0.346</td> <td>72.9</td> <td>60.1</td> <td>0.192</td> <td>0.35</td> <td>0.160</td> <td>0.139</td> <td>0.984</td> <td>0.433</td> </tr> <tr> <td>dots.ocr</td> <td>0.186</td> <td>0.198</td> <td><ins>0.115</ins></td> <td>0.169</td> <td>0.291</td> <td>0.358</td> <td>79.5</td> <td>82.5</td> <td>0.172</td> <td>0.141</td> <td>0.165</td> <td>0.123</td> <td>1.0</td> <td><ins>0.255</ins></td> </tr> <tr> <td>MonkeyOcr</td> <td>0.193</td> <td>0.259</td> <td>0.127</td> <td>0.236</td> <td>0.262</td> <td>0.325</td> <td>78.4</td> <td>74.7</td> <td>0.186</td> <td>0.294</td> <td>0.197</td> <td>0.180</td> <td>1.0</td> <td>0.623</td> </tr> <tr> <td>OCRFlux</td> <td>0.252</td> <td>0.254</td> <td>0.134</td> <td>0.195</td> <td>0.326</td> <td>0.405</td> <td>58.3</td> <td>70.2</td> <td>0.358</td> <td>0.260</td> <td>0.191</td> <td>0.156</td> <td>1.0</td> <td>0.284</td> </tr> <tr> <td>Gotocr</td> <td>0.247</td> <td>0.249</td> <td>0.181</td> <td>0.213</td> <td>0.231</td> <td>0.318</td> <td>59.5</td> <td>74.7</td> <td>0.38</td> <td>0.299</td> <td>0.195</td> <td>0.164</td> <td>0.969</td> <td>0.446</td> </tr> <tr> <td>Olmocr</td> <td>0.341</td> <td>0.382</td> <td>0.125</td> <td>0.205</td> <td>0.719</td> <td>0.766</td> <td>57.1</td> <td>56.6</td> <td>0.327</td> <td>0.389</td> <td>0.191</td> <td>0.169</td> <td>1.0</td> <td>0.294</td> </tr> <tr> <td>SmolDocling</td> <td>0.657</td> <td>0.895</td> <td>0.486</td> <td>0.932</td> <td>0.859</td> <td>0.972</td> <td>18.5</td> <td>1.5</td> <td>0.86</td> <td>0.98</td> <td>0.413</td> <td>0.695</td> <td>1.0</td> <td>0.927</td> </tr> <tr> <td>Logics-Parsing</td> <td>0.124</td> <td>0.145</td> <td>0.089</td> <td>0.139</td> <td><ins>0.106</ins></td> <td><ins>0.165</ins></td> <td>76.6</td> <td>79.5</td> <td>0.165</td> <td>0.166</td> <td><ins>0.136</ins></td> <td><ins>0.113</ins></td> <td>0.519</td> <td>0.252</td> </tr> <tr> <td rowspan="5">General VLMs</td> <td>Qwen2VL-72B</td> <td>0.298</td> <td>0.342</td> <td>0.142</td> <td>0.244</td> <td>0.431</td> <td>0.363</td> <td>64.2</td> <td>55.5</td> <td>0.425</td> <td>0.581</td> <td>0.193</td> <td>0.182</td> <td>0.792</td> <td>0.359</td> </tr> <tr> <td>Qwen2.5VL-72B</td> <td>0.233</td> <td>0.263</td> <td>0.162</td> <td>0.24</td> <td>0.251</td> <td>0.257</td> <td>69.6</td> <td>67</td> <td>0.313</td> <td>0.353</td> <td>0.205</td> <td>0.204</td> <td>0.597</td> <td>0.349</td> </tr> <tr> <td>Doubao-1.6</td> <td>0.188</td> <td>0.248</td> <td>0.129</td> <td>0.219</td> <td>0.273</td> <td>0.336</td> <td>74.9</td> <td>69.7</td> <td>0.180</td> <td>0.288</td> <td>0.171</td> <td>0.148</td> <td>0.601</td> <td>0.317</td> </tr> <tr> <td>GPT-5</td> <td>0.242</td> <td>0.373</td> <td>0.119</td> <td>0.36</td> <td>0.398</td> <td>0.456</td> <td>67.9</td> <td>55.8</td> <td>0.26</td> <td>0.397</td> <td>0.191</td> <td>0.28</td> <td>0.88</td> <td>0.46</td> </tr> <tr> <td>Gemini2.5 pro</td> <td>0.185</td> <td>0.20</td> <td><ins>0.115</ins></td> <td>0.155</td> <td>0.288</td> <td>0.326</td> <td><ins>82.6</ins></td> <td>80.3</td> <td>0.154</td> <td>0.182</td> <td>0.181</td> <td>0.136</td> <td><ins>0.535</ins></td> <td>0.26</td> </tr>

</table>  <tr> <td colspan="5"> * Tested on the v3/PDF Conversion API (August 2025 deployment).

</td> </tr>

Quick Start

1. Installation

shell

conda create -n logis-parsing python=3.10
conda activate logis-parsing

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124

2. Download Model Weights

# Download our model from Modelscope.
pip install modelscope
python download_model.py -t modelscope

# Download our model from huggingface.
pip install huggingface_hub
python download_model.py -t huggingface

3. Inference

shell

python3 inference.py --image_path PATH_TO_INPUT_IMG --output_path PATH_TO_OUTPUT --model_path PATH_TO_MODEL

Acknowledgments

We would like to acknowledge the following open-source projects that provided inspiration and reference for this work: