LightOnOCR-2-1B - An OCR model launched by LightOnAI

What is LightOnOCR-2-1B?

LightOnOCR-2-1B is a high-efficiency OCR model from LightOnAI, with only 1B of parameters. It excels in complex document processing, particularly in recognizing academic papers, mathematical formulas, and complex tables. Employing RLVR reinforcement learning, the model outputs structured text in Markdown format, supporting table reconstruction, formula recognition, and multi-column layout processing. Highly regarded for its low cost and high-speed processing capabilities, the model costs less than $0.01 per thousand pages of documents and achieves a processing speed of up to 5.71 pages per second, making it an ideal choice for document digitization and knowledge base construction.

Main functions of LightOnOCR-2-1B

• Markdown structured outputIt outputs the recognized text in Markdown format, including structured content such as headings, lists, and code blocks, making it easier to edit and format later.
• Table restorationIt can accurately restore the table structure and supports the recognition of complex statistical tables.
• Mathematical formula recognitionIt perfectly supports the recognition of mathematical formulas in LaTeX/KaTeX formats, making it suitable for academic literature and engineering documents.
• Multi-column layout processingAutomatically recognizes the reading order of multi-column layouts, suitable for complex document layouts such as newspapers and academic papers.
• Bounding box prediction (bbox variant)The model can recognize text and predict the bounding boxes of text in images, which facilitates image-text correspondence and slice processing.
• High efficiency and low costOn a single H100 graphics card, it can process 5.71 pages of documents per second, with a processing cost of less than $0.01 per thousand pages, making it suitable for large-scale document processing.

Technical Principles of LightOnOCR-2-1B

• Reinforcement Learning Based on Validation Feedback (RLVR):KaTeX's reward mechanism is optimized for rendering mathematical formulas, ensuring that the output LaTeX code is standardized and renderable.The reward mechanism is compressed to punish repetitive behaviors of the model, reducing the repetition rate by more than 50% and avoiding the dead loop problem common in small models.
• End-to-end OCR model architecture:supportInput is PDF or image, output is structured text directly, without the need for complicated preprocessing or postprocessing.
• High-efficiency reasoning framework:By combining the vLLM inference framework, the throughput and cost of the model are optimized, improving processing speed and economy.
• High-quality data cleaning:Use high-quality training data to ensure the accuracy and robustness of the model in complex scenarios.

The project address for LightOnOCR-2-1B

• HuggingFace model libraryhttps://huggingface.co/lightonai/LightOnOCR-2-1B
• arXiv technical paper: https://arxiv.org/pdf/2601.14251

Application scenarios of LightOnOCR-2-1B

• Digitalization of academic papers and arXiv documentsThe model can quickly convert complex formatting, mathematical formulas, and multi-column layouts in academic papers into structured Markdown format, making it easier for researchers and institutions to organize literature and share knowledge.
• OCR processing of old archives and scanned booksThe model can efficiently recognize text and tables in old documents, supporting the digital preservation and research of historical documents, and is suitable for institutions such as libraries and archives.
• Enterprise document platform and knowledge base constructionStructure internal documents, reports, and statements to facilitate knowledge management and data cleaning, thereby improving the efficiency of enterprise information systems.
• Financial notes and complex financial statement structuringThe model can quickly recognize text and tables in financial documents, enabling automated report processing and improving the efficiency and accuracy of financial data entry.
• Construction of Mathematics, Engineering and Scientific Knowledge BasesIt accurately identifies mathematical formulas and engineering drawings, supporting research institutions and universities in building high-quality knowledge bases to facilitate academic research and teaching.