POINTS-Reader: Distillation-Free Adaptation of Vision-Language Models for Document Conversion

HuggingFace arXiv view

We are delighted to announce that the WePOINTS family has welcomed a new member: POINTS-Reader, a vision-language model for end-to-end document conversion. ## News - 2025.08.26: We released the weights of the most recent version of POINT-ReaderπŸ”₯πŸ”₯πŸ”₯. - 2025.08.21: POINTS-Reader is accepted by **EMNLP 2025** for presentation at the **Main Conference**πŸŽ‰πŸŽ‰πŸŽ‰. ## Introduction 1. **Simplicity**: POINTS-Reader is a very streamlined model that fully follows the structure of POINTS1.5, except that we have replaced Qwen2.5-7B-Instruct with Qwen2.5-3B-Instruct. Moreover, the input and output of POINTS-Reader are extremely straightforward. The input consists of a fixed prompt and a document image, and the output contains only a string (text extracted from the document image). The model's output is the final result delivered to the user without any post-processing. 2. **Performance**: Currently, POINTS-Reader supports extraction from both Chinese and English documents, achieving impressive results, with scores of 0.133 for English and 0.212 for Chinese on OmniDocBench. 3. **High Throughput**: With current mainstream inference frameworks, such as SGLang and vLLM, optimization is predominantly focused on LLMs. Thus, a large ViT would significantly impact the model’s throughput, which is why we selected a ViT with a moderate number of parameters (600M NaViT used in POINTS1.5). Combined with our support for SGLang, we currently achieve a very satisfactory throughput. We will also provide support for vLLM in the future. 4. **Open-source Technical Approach**: In the POINTS-Reader paper, we propose a two-stage data augmentation strategy. The first stage leverages automated data to endow the model with basic document extraction capabilities. In the subsequent stage, continuous self-evolution improves the quality of data generated by the model. The self-evolution approach in the second stage is highly extensible and can be applied to virtually any model. ## Results We take the following results from [OmniDocBench](https://github.com/opendatalab/OmniDocBench/tree/main) and POINTS-Reader for comparison:
Method Type Methods OverallEdit↓ TextEdit↓ FormulaEdit↓ FormulaCDM↑ TableTEDS↑ TableEdit↓ Read OrderEdit↓
EN ZH EN ZH EN ZH EN ZH EN ZH EN ZH EN ZH
Pipeline Tools MinerU-pipeline-2.1.1 0.162 0.244 0.072 0.111 0.313 0.581 79.2 48.8 77.4 79.5 0.166 0.15 0.097 0.136
Marker-1.2.3 0.336 0.556 0.08 0.315 0.53 0.883 17.6 11.7 67.6 49.2 0.619 0.685 0.114 0.34
Marker-1.7.1 0.296 0.497 0.085 0.293 0.374 0.688 79.0 36.7 67.6 54.0 0.609 0.678 0.116 0.329
PaddleOCR PP-StructureV3 0.145 0.206 0.058 0.088 0.295 0.535 81.8 52.1 77.2 83.9 0.159 0.109 0.069 0.091
Mathpix 0.191 0.364 0.105 0.381 0.306 0.454 82.7 64.6 77.0 67.1 0.243 0.32 0.108 0.304
Docling-2.14.0 0.589 0.909 0.416 0.987 0.999 1 - - 61.3 25.0 0.627 0.810 0.313 0.837
Pix2Text-1.1.2.3 0.32 0.528 0.138 0.356 0.276 0.611 78.4 39.6 73.6 66.2 0.584 0.645 0.281 0.499
Unstructured-0.17.2 0.586 0.716 0.198 0.481 0.999 1 - - 0 0.064 1 0.998 0.145 0.387
OpenParse-0.7.0 0.646 0.814 0.681 0.974 0.996 1 0.106 0 64.8 27.5 0.284 0.639 0.595 0.641
Expert VLMs POINTS-Reader-3B 0.133 0.212 0.062 0.139 0.304 0.465 - - 83.7 85.0 0.128 0.136 0.036 0.106
MinerU2.0-2505-0.9B 0.133 0.238 0.045 0.115 0.273 0.506 79.0 50.8 82.1 83.4 0.15 0.209 0.066 0.122
MonkeyOCR-pro-1.2B 0.146 0.221 0.068 0.118 0.272 0.452 76.7 63.3 81.3 85.5 0.149 0.134 0.093 0.179
Dolphin 0.356 0.440 0.352 0.440 0.465 0.604 61.6 40.4 70.2 56.8 0.258 0.367 0.35 0.351
Nanonets-OCR-s 0.283 0.295 0.134 0.231 0.518 0.546 63.2 52.0 76.8 79.4 0.343 0.201 0.135 0.2
OCRFlux-3B 0.238 0.349 0.112 0.256 0.447 0.716 60.2 31.9 69.0 80.0 0.269 0.162 0.126 0.263
GOT-OCR 0.287 0.411 0.189 0.315 0.360 0.528 74.3 45.3 53.2 47.2 0.459 0.52 0.141 0.28
Nougat 0.452 0.973 0.365 0.998 0.488 0.941 15.1 16.8 39.9 0.0 0.572 1.000 0.382 0.954
Mistral OCR 0.268 0.439 0.072 0.325 0.318 0.495 64.6 45.9 75.8 63.6 0.6 0.65 0.083 0.284
OLMOCR-sglang 0.326 0.469 0.097 0.293 0.455 0.655 74.3 43.2 68.1 61.3 0.608 0.652 0.145 0.277
SmolDocling-256M_transformer 0.493 0.816 0.262 0.838 0.753 0.997 32.1 0.551 44.9 16.5 0.729 0.907 0.227 0.522
General VLMs
Gemini2.0-flash 0.191 0.264 0.091 0.139 0.389 0.584 77.6 43.6 79.7 78.9 0.193 0.206 0.092 0.128
Gemini2.5-Pro 0.148 0.212 0.055 0.168 0.356 0.439 80.0 69.4 85.8 86.4 0.13 0.119 0.049 0.121
GPT4o 0.233 0.399 0.144 0.409 0.425 0.606 72.8 42.8 72.0 62.9 0.234 0.329 0.128 0.251
Qwen2-VL-72B 0.252 0.327 0.096 0.218 0.404 0.487 82.2 61.2 76.8 76.4 0.387 0.408 0.119 0.193
Qwen2.5-VL-7B 0.316 0.399 0.151 0.243 0.376 0.5 75.3 57.3 71.1 71.3 0.598 0.627 0.138 0.226
Qwen2.5-VL-72B 0.214 0.261 0.092 0.18 0.315 0.434 81.4 64.1 81.4 83.0 0.341 0.262 0.106 0.168
InternVL2-76B 0.44 0.443 0.353 0.290 0.543 0.701 67.4 44.1 63.0 60.2 0.547 0.555 0.317 0.228
InternVL3-78B 0.218 0.296 0.117 0.21 0.38 0.533 79.2 58.8 69.0 73.9 0.279 0.282 0.095 0.161
## Getting Started This following code snippet has been tested with following environment: ``` python==3.10.12 torch==2.5.1 transformers==4.46.1 cuda==12.1 ``` If you encounter environment issues, please feel free to open an issue. ### Run with Transformers Before you run the following code, make sure you install the `WePOINTS` package by running: ``` git clone https://github.com/WePOINTS/WePOINTS.git cd WePOINTS pip install -e. ``` ```python from wepoints.utils.images import Qwen2ImageProcessorForPOINTSV15 from transformers import AutoModelForCausalLM, AutoTokenizer import torch # We recommend using the following prompt to better performance, # since it is used throughout the training process. prompt = ( 'Please extract all the text from the image with the following requirements:\n' '1. Return tables in HTML format.\n' '2. Return all other text in Markdown format.' ) image_path = '/path/to/your/local/image' model_path = 'tencent/POINTS-Reader' model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype=torch.float16, device_map='cuda') tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) image_processor = Qwen2ImageProcessorForPOINTSV15.from_pretrained(model_path) content = [ dict(type='image', image=image_path), dict(type='text', text=prompt) ] messages = [ { 'role': 'user', 'content': content } ] generation_config = { 'max_new_tokens': 2048, 'repetition_penalty': 1.05, 'temperature': 0.7, 'top_p': 0.8, 'top_k': 20, 'do_sample': True } response = model.chat( messages, tokenizer, image_processor, generation_config ) print(response) ``` If you encounter issues like repeation, please try to increase the resolution of the image to allievate the problem. ### Deploy with SGLang We will create a Pull Request to SGLang, please stay tuned. ## Known Issues - **Complex Document Parsing**: POINTS-Reader can struggle with complex layouts (e.g., newspapers), often producing repeated or missing content. - **Handwritten Document Parsing**: It also has difficulty handling handwritten inputs (e.g., receipts, notes), which can lead to recognition errors or omissions. - **Multi-language Document Parsing**: POINTS-Reader currently supports only English and Chinese, limiting its effectiveness on other languages. ## Citation If you use this model in your work, please cite the following paper: ``` @article{points-reader, title={POINTS-Reader: Distillation-Free Adaptation of Vision-Language Models for Document Conversion}, author={Liu, Yuan and Zhongyin Zhao and Tian, Le and Haicheng Wang and Xubing Ye and Yangxiu You and Zilin Yu and Chuhan Wu and Zhou, Xiao and Yu, Yang and Zhou, Jie}, journal={}, year={2025} } @article{liu2024points1, title={POINTS1. 5: Building a Vision-Language Model towards Real World Applications}, author={Liu, Yuan and Tian, Le and Zhou, Xiao and Gao, Xinyu and Yu, Kavio and Yu, Yang and Zhou, Jie}, journal={arXiv preprint arXiv:2412.08443}, year={2024} } @article{liu2024points, title={POINTS: Improving Your Vision-language Model with Affordable Strategies}, author={Liu, Yuan and Zhao, Zhongyin and Zhuang, Ziyuan and Tian, Le and Zhou, Xiao and Zhou, Jie}, journal={arXiv preprint arXiv:2409.04828}, year={2024} } @article{liu2024rethinking, title={Rethinking Overlooked Aspects in Vision-Language Models}, author={Liu, Yuan and Tian, Le and Zhou, Xiao and Zhou, Jie}, journal={arXiv preprint arXiv:2405.11850}, year={2024} } ```