What is LightOnOCR-2-1B
LightOnOCR-2-1B is an efficient OCR model released by LightOnAI, with only 1B parameters, delivering outstanding performance in complex document processing and excelling at recognizing academic papers, mathematical formulas, and complex tables. The model uses RLVR reinforcement learning technology to output structured text in Markdown format, supporting table reconstruction, formula recognition, and multi-column layout processing. The model stands out for its low cost and high-speed processing capability, with a processing cost of less than $0.01 per thousand pages and a processing speed of up to 5.71 pages/second, making it an ideal choice for document digitization and knowledge base construction.
Main features of LightOnOCR-2-1B
- Markdown structured output: Outputs the recognized text in Markdown format, including structured content such as headings, lists, and code blocks, making subsequent editing and formatting easier.
- Table reconstruction: Can accurately reconstruct table structures and supports recognition of complex statistical tables.
- Mathematical formula recognition: Fully supports recognition of mathematical formulas in LaTeX/KaTeX format, suitable for academic literature and engineering documents.
- Multi-column layout processing: Automatically recognizes the reading order of multi-column layouts, suitable for complex formatted documents such as newspapers and academic papers.
- Bounding box prediction (bbox variant): The model can recognize text and predict bounding boxes of text in images, facilitating image-text correspondence and slicing.
- High efficiency and low cost: On a single H100 GPU, it can process 5.71 pages per second, with a processing cost of less than $0.01 per thousand pages, suitable for large-scale document processing.
Technical principles of LightOnOCR-2-1B
-
Reinforcement Learning with Verification Feedback (RLVR)：The KaTeX reward mechanism is optimized for mathematical formula rendering, ensuring that the output LaTeX code is standardized and renderable.The compression reward mechanism penalizes repeated behavior in the model, reducing the repetition rate by more than 50% and avoiding the infinite loop problem common in small models.
-
End-to-end OCR model architecture: Supportsinput as PDF or image, directly outputting structured text without complex preprocessing or post-processing workflows.
-
Efficient inference framework：Combined with the vLLM inference framework, it optimizes the model's throughput and cost, improving processing speed and cost-effectiveness.
-
High-quality data cleaning：Uses high-quality training data to ensure the model's accuracy and robustness in complex scenarios.
Project地址 of LightOnOCR-2-1B
- HuggingFace model repository：https://huggingface.co/lightonai/LightOnOCR-2-1B
- arXiv technical paper：https://arxiv.org/pdf/2601.14251
Application scenarios of LightOnOCR-2-1B
-
Digitization of academic papers and arXiv literature: The model can quickly convert complex layouts, mathematical formulas, and multi-column layouts in academic papers into structured Markdown format, facilitating literature organization and knowledge sharing for researchers and institutions.
-
OCR processing of old archives and scanned books: The model can efficiently recognize text and tables in old documents, supporting digital preservation and research of historical literature, suitable for institutions such as libraries and archives.
-
Enterprise document middle platform and knowledge base construction: Structurally processes internal enterprise documents, reports, and statements, facilitating knowledge management and data cleaning, and improving the efficiency of enterprise information systems.
-
Structuring of financial receipts and complex statements: The model can quickly recognize text and tables in financial documents, enabling automated processing of reports and improving the efficiency and accuracy of financial data entry.
-
Construction of mathematics, engineering, and scientific research knowledge bases: It accurately recognizes mathematical formulas and engineering drawings, supporting research institutions and universities in building high-quality knowledge bases for academic research and teaching.