Description
A scanned paper with equations, tables and two columns crammed together: copy the text out and you get garbage or scrambled order, and cleaning it up for an LLM dataset or a knowledge base eats half a day. olmOCR, an open-source toolkit from Ai2, uses a 7B-parameter vision-language model to turn PDFs, PNGs and JPEGs straight into clean Markdown, laid out in natural reading order, even around figures, multi-column pages and insets.
It puts effort into the hard cases: equations, tables, handwriting and complex layouts are handled, and headers and footers are removed automatically. The project quotes under $200 per million pages, which suits large batches, and ships olmOCR-Bench, a benchmark of more than 7,000 test cases over 1,400 documents for comparing OCR systems.
Several inputs: PDF, PNG and JPEG documents all convert to one Markdown output.
Hard content: equations, tables, handwriting and complex formatting are recognised, and headers and footers are stripped out of the body text.
Reading order: text stays in the order a person would read it, with figures, columns and insets in the way.
Batch and cluster runs: vLLM-based inference on a single GPU, an external inference server, or a multi-node S3 work queue for large PDF collections, plus official Docker images.
Benchmark and training: olmOCR-Bench is included and the training code is open, so you can train your own olmOCR model.
Requirements: built on a 7B vision-language model, so an NVIDIA GPU is required; a Python package under Apache-2.0, with an online demo to try first.
It puts effort into the hard cases: equations, tables, handwriting and complex layouts are handled, and headers and footers are removed automatically. The project quotes under $200 per million pages, which suits large batches, and ships olmOCR-Bench, a benchmark of more than 7,000 test cases over 1,400 documents for comparing OCR systems.
Features
Several inputs: PDF, PNG and JPEG documents all convert to one Markdown output.
Hard content: equations, tables, handwriting and complex formatting are recognised, and headers and footers are stripped out of the body text.
Reading order: text stays in the order a person would read it, with figures, columns and insets in the way.
Batch and cluster runs: vLLM-based inference on a single GPU, an external inference server, or a multi-node S3 work queue for large PDF collections, plus official Docker images.
Benchmark and training: olmOCR-Bench is included and the training code is open, so you can train your own olmOCR model.
Requirements: built on a 7B vision-language model, so an NVIDIA GPU is required; a Python package under Apache-2.0, with an online demo to try first.
