allenai/olmOCR-2-7B-1025
The allenai/olmOCR-2-7B-1025 is a 7 billion parameter vision-language model developed by AllenAI, fine-tuned from Qwen2.5-VL-7B-Instruct. It specializes in Optical Character Recognition (OCR) for complex document layouts, including math equations and tables, leveraging additional fine-tuning with GRPO RL training. This model is designed for high-accuracy text extraction from PDFs, particularly challenging scans and multi-column documents, and is best utilized with the olmOCR toolkit for efficient inference.
Loading preview...
olmOCR-2-7B-1025 Overview
olmOCR-2-7B-1025 is a 7 billion parameter vision-language model developed by AllenAI, fine-tuned from Qwen2.5-VL-7B-Instruct. It is specifically designed for advanced Optical Character Recognition (OCR) tasks, excelling at extracting text from complex document structures like math equations, tables, and challenging scanned documents. The model underwent additional fine-tuning using GRPO RL training on the olmOCR-mix-1025 and olmOCR-synthmix-1025 datasets to boost its performance in these tricky OCR cases.
Key Capabilities
- High-Accuracy OCR: Specialized in extracting text from diverse and challenging PDF documents.
- Complex Layout Handling: Demonstrates strong performance on math equations, tables, old scans, and multi-column layouts.
- Vision-Language Integration: Leverages a vision-language architecture for robust document understanding.
- Optimized for Toolkit Use: Best utilized with the olmOCR toolkit for efficient inference, automatic rendering, rotation, and retries.
Good For
- Document Digitization: Ideal for converting scanned PDFs, academic papers (ArXiv), and historical documents into searchable text.
- Data Extraction: Excellent for extracting structured data from tables and complex mathematical expressions.
- Research and Development: Suitable for researchers and developers working on advanced document understanding and information retrieval from PDFs.
- Large-Scale Processing: The accompanying olmOCR toolkit supports efficient inference for millions of documents at scale.