allenai/olmocr is a collection of tools that converts documents delivered as screen images — such as PDF, PNG and JPEG — into Markdown and readable plain text. The repository describes this purpose as a document linearization (linearizing PDFs) toolkit, and its output is used for LLM datasets and training. The latest GitHub release is v0.4.27.
The README explains that it handles formulas, tables, handwriting and complex formatting, removes headers and footers, and converts to a natural reading order even with figures or multi-column layouts. It is the olmOCR model pipeline that processes body text, tables, formulas and reading order in a single flow. The implementation language is Python and the license is Apache-2.0, with 19776 GitHub stars. The repository was created on September 17, 2024, and because the model is based on a 7B VLM (vision-language model), a GPU is required. It also provides olmOCR-Bench, which measures the performance of OCR systems.
That said, since this record has no changelog, there is no way to know what changed in v0.4.27. olmOCR-Bench is said to be provided, but this record also contains no olmOCR-Bench scores and no comparison results against other models. The mere presence of an evaluation tool does not mean good results came out of it. The only figure left in the record is a momentum score of 5.1 (out of 10), while the technical, quality and originality scores are empty. Because the 7B model requires a GPU, throughput and cost are points to check separately. As performance evidence is still missing, it is worth reviewing as a candidate for document preprocessing in a training pipeline.