This toolkit addresses the challenge of extracting and linearizing text from diverse PDF documents for large language model (LLM) training datasets. It offers a comprehensive pipeline for processing millions of PDFs, targeting researchers and engineers working with document understanding and LLM fine-tuning.

How It Works

The core of olmOCR leverages a prompting strategy with GPT-4o for high-quality text parsing from PDF images. It processes PDFs through a fine-tuned Qwen2-VL and Molmo-O model, utilizing Sglang for efficient batch inference. The output is structured in Dolma-style JSONL, facilitating easy integration into LLM training workflows.

Quick Start & Requirements

• Installation: pip install -e .[gpu] --find-links https://flashinfer.ai/whl/cu124/torch2.4/flashinfer/
• Prerequisites: NVIDIA GPU (>=20GB VRAM), 30GB disk space, poppler-utils, and Microsoft Core Fonts.
• Setup: Requires conda environment setup and installing dependencies.
• Demo: Online demo available at https://olmocr.allenai.org/.

Highlighted Details

• Utilizes a GPT-4o prompting strategy for robust PDF text extraction.
• Supports batch processing of millions of PDFs via S3 integration and cluster execution (e.g., Beaker).
• Includes fine-tuning code for Qwen2-VL and Molmo-O models.
• Offers a side-by-side evaluation toolkit for comparing pipeline versions.

Maintenance & Community

Developed and maintained by the AllenNLP team at the Allen Institute for Artificial Intelligence (AI2).

Licensing & Compatibility

Licensed under Apache 2.0, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

Requires a recent NVIDIA GPU with substantial VRAM for inference. The project is presented with a 2025 arXiv publication date, suggesting it may be relatively new or undergoing active development.