From the source
Ai2 is documenting how two research teams adapted its open olmOCR 2 model to historical and archaeological archives, running it locally rather than through commercial APIs.
- Historian Maurice Casey built ArtificeOCR, which connects locally running vision-language models such as olmOCR 2 to Tropy, the open-source application he uses to organize archival photographs, making his library searchable while preserving metadata entered since 2019.
- At TU Wien's Computer Vision Lab and the Austrian Archaeological Institute, researchers fine-tuned olmOCR 2 with LoRA on 57 annotated records from a Carnuntum pottery archive of roughly 70,000 records, and transcription error on the remaining 450 test records fell from 15.58% to 1.11%.
- Of the five open models tested on the team's 507-record CENTURIA benchmark, olmOCR 2 had the lowest transcription error before adaptation.
- Casey reports strong results with cursive English and typescript, but also errors when the model transcribes handwriting in languages less represented in its training data, and ArtificeOCR is still in early development.
- The team's next goal is to apply the approach to the full roughly 70,000 pottery records, which the researchers say would make comparisons across the site and between periods tractable.
