Sarashina2.2 Ocr

--- language: - ja - en base_model: - sbintuitions/sarashina2.2-3b-instruct-v0.1 license: mit tags: - multimodal - ocr - document-understanding - vision-language pipeline_tag: image-to-text library_name: transformers --- <div align="center"> 　<h1>Sarashina2.2-OCR: <br>End-to-end OCR Model for Japanese Document Parsing</h1> <a href="https://www.sbintuitions.co.jp/en/" target="_blank"> <img alt="Website" src="https://img.shields.io/badge/Website-SB%20Intuitions-777777?labelColor=555555" /> </a> <a href="https://huggingface.co/sbintuitions" target="_blank"> <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-SB%20Intuitions-777777?labelColor=555555" /> </a> <a href="https://x.com/sbintuitions" target="_blank"> <img alt="Twitter Follow" src="https://img.shields.io/badge/X-@sbintuitions-777777?labelColor=555555&logo=x&logoColor=white" /> </a> </div> ### **About the model** **Sarashina2.2-OCR** is an end-to-end 3B-parameter OCR model developed by **SB Intuitions**, specifically tailored for parsing Japanese and English documents. The model is refined by human preference optimization to enable intuitive document parsing, and excels at converting a wide range of documents—including vertical Japanese text—into Markdown format while maintaining a natural reading order. ### **Key Features** Beyond standard text extraction, Sarashina2.2-OCR reconstructs documents into naturally structured Markdown, accurately converting complex elements into the following formats: - 📊 **Tables**: Reconstructs tabular layouts into plain **HTML** format. - 📐 **Math Formulas**: Transcribes mathematical equations directly into standard **LaTeX** format. - 🖼️ **Graphics**: Detects visual components (e.g., images, charts) and indicates their positions using **bounding boxes** in the format &lt;bbox&gt;[(x1, y1), (x2, y2)]&lt;/bbox&gt;, using normalized integer coordinates (0–1000) with a top-left origin. ## **Training Summary** Sarashina2.2-OCR integrates a [SigLIP2](https://huggingface.co/google/siglip2-so400m-patch14-384)-based vision encoder with the [Sarashina2.2-3B-Instruct](https://huggingface.co/sbintuitions/sarashina2.2-3b-instruct-v0.1) language model, and was trained through the following pipeline: ### **1. Pre-training on general image-text dataset:** To build basic image understanding and prepare for high-resolut...

Source context: 1617 downloads · 33 likes · Pipeline image-to-text · Library transformers · Repo sbintuitions/sarashina2.2-ocr