olmOCR

olmOCR converts scanned and image-based documents into text for researchers and developers preparing LLM datasets. It handles pages whose layout makes ordinary text extraction difficult.

• Document conversion: Convert PDF, PNG and JPEG documents into Markdown, with Dolma output for dataset workflows.
• Complex page content: Extract equations, tables and handwriting alongside text in heavily formatted documents.
• Reading order: Keep text in a natural sequence across columns, figures and insets, and remove page headers and footers automatically.
• Large collections: Spread processing across multiple worker nodes, with AWS S3 support for document input and work coordination.
• OCR evaluation: Use olmOCR-Bench to compare OCR systems on difficult document content and layouts.

The Python toolkit runs on your own GPU and is available in Docker images. Local inference requires a GPU. It can also use vLLM or another server implementing the OpenAI API; remote inference sends document content to that server. Cirrascale is a verified external provider, and a hosted web demo is available. The toolkit is open source under Apache 2.0.