Olmocr 2: Unit Test Rewards for Document Ocr

(From Ai2's Discord channel)

We’re rolling out *olmOCR 2*—the next major update to our open OCR model for complex documents & scans.

olmOCR 2 turns messy files with tables, equations, handwriting, and more into clean text. Under the hood, we combine synthetic data with unit tests as verifiable rewards to push state-of-the-art performance on challenging docs.

*What’s new*

◆ *Stronger text recognition:* Trained with a new data mix, including 20,000 historical pages for better coverage of aged and degraded materials. Example: olmOCR 2 can now read Abraham Lincoln’s handwriting correctly, recovering the date “January 10th” in his 1864 letter to Major General Hitchcock.

◆ *Big benchmark gains:* 82.4 on olmOCR-Bench (up from 78.5), with improvements across every document category.

◆ *Faster & cheaper:* New FP8 quantized model (olmOCR-2-7B-1025-FP8) reaches ~3,400 output tokens/sec on a single H100—enough to process 10,000 pages for < $2.

◆ *Adapt to your data:* Want to fine-tune for your domain? We provide everything you need to customize and deploy.

Available now, and on the DeepInfra & Parasail APIs. We’re also updating our demo—try olmOCR 2 today!

Learn more: https://allenai.org/blog/olmocr-2 Model: https://huggingface.co/allenai/olmOCR-2-7B-1025-FP8