olmOCR 2 cut pottery-record transcription errors from 15.58% to 1.11% after training on 57 examples

olmOCR 2, an open model from Ai2 that reads text from images of documents, is being used in two archive projects. Ai2 describes both in a blog post. One is about privacy. The other is about accuracy.

A historian keeps his archive at home

Maurice Casey studies radical politics in early twentieth-century Europe. His sources are in English, Russian and German. He has thousands of photos of letters, newspapers and other records, taken on different smartphones in different conditions.

He wanted to search them by keyword. He did not want to upload material that archives and families had trusted him with to an outside service.

So he built ArtificeOCR. It links locally running models such as olmOCR 2 to Tropy, the open-source app he uses to organise archival photos. His library became searchable, and the notes he had entered since 2019 stayed intact.

Casey says the model copes with varied layouts, including the standard forms found in secret police archives, without special preparation for each one. He also reports strong results on cursive English and typescript.

Fluent text can hide mistakes

Old-style OCR, the software that turns a picture of text into words, tends to fail in obvious ways: garbled output. Casey notes that a language model can instead produce text that reads smoothly but is wrong. He also saw errors on handwriting in languages the model has seen less of.

ArtificeOCR lets a researcher put a scan next to its transcription, fix mistakes and review later changes.

Roman pottery notes

A team from the Computer Vision Lab at TU Wien and the Austrian Archaeological Institute looked at handwritten pottery records from Carnuntum, a Roman site in modern Austria. The archive holds roughly 70,000 records. The notes use abbreviated German, sit around or over technical drawings, and vary by decade.

The team built CENTURIA, a benchmark of 507 annotated records. Of five open models, olmOCR 2 had the lowest error before any adaptation. Using LoRA, which trains a small set of extra parameters, they tuned it on 57 records. On the other 450, transcription error fell from 15.58% to 1.11%.

Imagine searching 70,000 paper notes by hand. That is the problem they want to remove.

The limits are clear in the source. ArtificeOCR is in early development. The pottery result comes from a 507-record benchmark, and applying it to the full 70,000 records is the team's next goal. Casey's accounts are his own experience, not a formal test.