olmOCR-2
Second-gen open document-OCR toolkit that linearizes PDFs using unit-test RL rewards.
About
olmOCR-2 is the Allen Institute for AI's 2025 follow-up to olmOCR, its open toolkit for turning PDFs into clean linear text for LLM datasets and training. Where v1 paired a vision-language model with the OlmOCR-bench evaluation set, v2 trains with reinforcement learning driven by unit-test rewards — the model is rewarded for OCR outputs that pass programmatic checks — aiming at more reliable document conversion. The paper is arXiv 2510.19817.
Like its predecessor it is Apache 2.0 and aimed squarely at the dataset builders who need trillions of PDF tokens converted faithfully: reading order, tables, math and layout preserved as structured text rather than the soup that naive PDF extraction produces. There is no hosted service; you run the pipeline yourself from the allenai/olmocr repository.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published benchmarks found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: arXiv paper (2510.19817) describing the unit-test RL method; Apache 2.0 toolkit on GitHub (allenai/olmocr); builds on the well-received olmOCR v1 and its OlmOCR-bench. Strengths: open license, credible lab, clear dataset-building use case. Docked for: (1) no benchmark figures found for v2 in sources checked; (2) self-host only, no managed option. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Paper arXiv 2510.19817 (2025): unit-test rewards RL for document OCR
- · Apache 2.0 toolkit (allenai/olmocr); builds on well-received olmOCR v1
- · No benchmark figures found for v2 in sources checked