ebinan92/Rukopys-OCR-4B
Rukopys-OCR-4B by ebinan92 is a vision-language model fine-tuned from Qwen3.5-4B, designed for Ukrainian handwritten document Optical Character Recognition (OCR). It specializes in detecting and classifying document regions, then transcribing their content into structured JSON, including handwritten, printed, formula, and table elements. This model was developed for the Handwritten to Data competition, where it contributed to a 3rd-place solution, excelling at extracting structured data from complex Ukrainian documents.
Loading preview...
Model Overview
Rukopys-OCR-4B is a specialized vision-language model developed by ebinan92, based on a full fine-tune of the Qwen3.5-4B architecture. Its primary function is to perform Optical Character Recognition (OCR) on Ukrainian handwritten documents, extracting structured information in JSON format.
Key Capabilities
- Region Detection and Classification: Identifies various regions within a document, including
handwritten,printed,formula,table,annotation,image, andgraph. - Structured Transcription: Transcribes text from detected regions, outputting a JSON array where each region includes its bounding box (
bbox),type, andtext(transcription). - Multimodal Input: Processes both image and text prompts to generate structured OCR output.
- Competition Proven: Formed part of the 3rd-place solution in the Handwritten to Data competition.
Training and Data
The model was trained using a diverse set of Ukrainian and Cyrillic handwriting datasets, including RUKOPYS, Ukrainian Handwritten Text, school_notebooks_RU, and HWR200. It also incorporated pseudo-labels generated with Gemini (gemini-3-flash-preview).
Good For
- Automated processing and digitization of Ukrainian handwritten documents.
- Extracting structured data from complex document layouts, including tables and formulas.
- Research and applications requiring high-accuracy OCR for historical or administrative Ukrainian texts.