> ## Content Index
> Fetch the complete content index at: https://soundbytes.news/llms.txt
> Use this file to discover other available public pages before exploring further.

# Open model adapted to read Roman pottery records and search photos, team says
- URL: https://soundbytes.news/open-model-adapted-to-read-roman-pottery-records-and-search-photos-team-says/
- Published: 2026-10-28T17:07:50.000Z
- Updated: 2026-11-18T08:49:34.000Z
- Description: Researchers in Austria adapted the Allen Institute for AI's open olmOCR 2 to read handwritten records, and a historian runs it on his own computer. The results are the team's own, reported in an Ai2 post.
- Author: Echo
- Tags: Open models

A team of researchers adapted an open text-reading model to read handwritten notes on Roman pottery records. Its transcription error on test records fell from 15.58% to 1.11%, the team says, in [an Allen Institute for AI (Ai2) blog post dated 28 October 2026](https://allenai.org/blog/olmocr-2-historical-archives?ref=soundbytes.news).

The model is olmOCR 2, Ai2's model for pulling searchable text out of document images, including handwriting. Open here means its weights (the numbers it has learned), fine-tuning data and code are available, so researchers can run it locally and adapt it.

The records come from Carnuntum, a Roman site in present-day Austria. Researchers at TU Wien and the Austrian Archaeological Institute study them in a project called LEGION.

The team built CENTURIA, a benchmark (a standard test set) of 507 annotated records. Of five open models tested, olmOCR 2 had the lowest transcription error before adaptation.

The researchers then fine-tuned it, which means training it further on a small set of examples. They used LoRA, a method that trains a small set of extra numbers, on 57 records, and measured the error on the other 450.

CENTURIA co-author Dominik Hagmann and colleagues write that they wanted a method heritage institutions could reproduce, as those often lack resources for commercial services.

The team's next goal is to apply it to the roughly 70,000 records in its archive, which the post says could help archaeologists compare pottery forms across the site and between historical periods.

Maurice Casey, a historian of radical politics in early twentieth-century Europe, built up thousands of smartphone photographs of letters, newspapers and other records. He wanted to search them by keyword without uploading material entrusted to him by archives and families to an outside service.

He built ArtificeOCR, which links models running on his own computer, such as olmOCR 2, with Tropy, the open-source application he uses to organise archive photographs. It kept the notes he had entered since 2019.

"The ability to run the model locally was the most important part of this for me," Casey says in the post.

Casey reports strong results with cursive English and typed pages across his research languages.

He also says a language model can produce fluent, plausible text where older software would return obvious garbling, and he reports errors on handwriting in languages less represented in its training data. In ArtificeOCR, researchers can compare a scan with its transcription and correct mistakes.

ArtificeOCR is still in early development. Casey wants interfaces that make local models easier for historians to use.

## Sources

- [Olmocr 2 historical archives](https://allenai.org/blog/olmocr-2-historical-archives?ref=soundbytes.news) · allenai.org