Researchers adapt Ai2's olmOCR 2 for historical and archaeological archives

An open model that runs locally

olmOCR 2 is Ai2's open model for extracting searchable, structured text from document images, including handwriting and complex layouts. Ai2 releases its weights, fine-tuning data and code, which lets researchers run it on their own hardware and adapt it to their needs.

Ai2 highlights two projects that use this openness in different ways: ArtificeOCR, built by historian Maurice Casey, and CENTURIA, developed by researchers working on Roman pottery records from Carnuntum, a site in present-day Austria.

ArtificeOCR: searching a private archive without uploading it

Casey studies radical politics in early twentieth-century Europe and works with sources in English, Russian and German. Over years of research he gathered thousands of photographs of letters, newspapers and other records, taken on different smartphones under varying conditions. He wanted to search them by keyword without uploading material entrusted to him by archives and families to a third-party service.

His tool, ArtificeOCR, connects locally running vision-language models such as olmOCR 2 with Tropy, the open-source application he uses to organize archival photographs. Casey made his existing Tropy library searchable while preserving the metadata he had entered since 2019.

According to Ai2, Casey finds the model especially useful for varied page layouts, including the standardized forms in secret police archives, without preparing each document format differently. He also reports strong results with cursive English and typescript across his research languages.

He points to limits as well. A language model can produce plausible, fluent text where conventional OCR would return obvious garbling, and he reports errors when the model transcribes handwriting in languages less represented in its training data. ArtificeOCR therefore lets researchers compare a scan with its transcription, correct mistakes and inspect changes made during later processing. "Fidelity to the source material, including capturing how people in the past made errors or wrote in coded or euphemistic language, is hugely important for my work as a historian," Casey says.

The tool is still in early development. Casey wants interfaces that make local models easier for historians to use and sees potential in fine-tuning them for particular kinds of sources.

CENTURIA: fine-tuning on Roman pottery records

Researchers at TU Wien's Computer Vision Lab and the Austrian Archaeological Institute at the Austrian Academy of Sciences studied handwritten pottery records from Carnuntum. Their archive holds roughly 70,000 records. Handwritten notes identify vessel forms, excavation contexts and other details archaeologists need to compare finds. The annotations appear around technical drawings, sometimes overlap them, and use abbreviated German terminology. Differences in handwriting and recording conventions across decades make transcription harder.

As part of the research project LEGION, the team created CENTURIA, a benchmark of 507 annotated records for evaluating transcription and extraction of structured information. Of the five open models tested, olmOCR 2 had the lowest transcription error before adaptation. Access to its weights allowed the team to train it on examples from the collection.

Using LoRA, a method that trains a small set of additional parameters, the researchers fine-tuned olmOCR 2 on 57 records. On the remaining 450 test records, transcription error fell from 15.58% to 1.11%.

The authors say they wanted a pipeline that is reproducible and realistically deployable by heritage institutions, "which often do not have the resources to process tens of thousands of documents through commercial APIs."

What it means for archives

The team's next goal is to apply the approach to the roughly 70,000 pottery records in the Carnuntum archive. Making the full collection searchable could help archaeologists compare pottery forms across the site and between historical periods. In the authors' words, such questions are "impractical when your data lives in paper annotations," but become "tractable" once records are structured and searchable.

Ai2 notes that archives differ in recording conventions and conditions of access, and that with the weights and code available, researchers can choose where to run olmOCR 2 and adapt it to their documents. Both projects build expert review into the process. For Casey, local operation matters most: "The ability to run the model locally was the most important part of this for me," he says. "There's so much material I simply would not feel comfortable sending to a third party."