olmOCR

olmOCR is an open-source document processing toolkit developed by the Allen Institute for AI (AI2) that converts PDFs and scanned document images into clean, structured plain text. It leverages vision-language models to accurately extract text from complex layouts, tables, and figures. Designed for large-scale data pipelines, olmOCR is optimized for processing millions of documents efficiently. It is particularly useful for researchers and engineers building training datasets for large language models.

• PDF & Image OCR: Convert PDFs and scanned images to plain text using state-of-the-art vision-language models for high accuracy on complex layouts.
• Large-Scale Processing: Built for throughput, olmOCR can handle millions of documents in batch pipelines, making it suitable for dataset construction at scale.
• Structured Text Output: Preserves document structure including headings, tables, and lists in the extracted text output.
• Open Source: Fully open-source under a permissive license, allowing researchers and developers to inspect, modify, and extend the codebase freely.
• CLI & Python API: Accessible via command-line interface and Python API, enabling easy integration into existing data processing workflows.
• Model-Backed Extraction: Uses AI2's OLMo-family vision-language models to power document understanding beyond simple character recognition.
• Batch Pipeline Support: Designed to integrate into distributed computing environments for processing large document corpora efficiently.
• Research-Grade Quality: Developed by AI2 researchers with a focus on producing high-quality text for LLM pre-training and academic research use cases.