Ai2's Open Model olmOCR 2 Reduces Archaeological Handwriting Transcription Errors from 15.58% to 1.11% After Fine-Tuning
Researchers used LoRA fine-tuning on Ai2's olmOCR 2 model to significantly improve transcription accuracy for handwritten archaeological records at Carnuntum, enabling local, privacy-preserving archive search.
Non-profit AI research institute Allen Institute for AI (Ai2) announced in a blog post that researchers are adapting its open-source document image extraction model, olmOCR 2, for historical and archaeological archives. The model supports local execution, allowing users to perform keyword searches and structured text extraction without uploading sensitive materials to third-party services.
In a collaboration between TU Wien’s Computer Vision Lab and the Austrian Archaeological Institute at the Austrian Academy of Sciences, part of the LEGION project, researchers tested the model on handwritten pottery records from the Roman site of Carnuntum in Austria. These records comprise approximately 70,000 documents, with annotations often overlapping technical drawings and using abbreviated German terminology. The team created the CENTURIA benchmark dataset containing 507 annotated records.
According to the Ai2 blog, before adaptation, olmOCR 2 had the lowest transcription error among five tested open models. The research team used Low-Rank Adaptation (LoRA) to fine-tune the model on just 57 samples. On the remaining 450 test records, the transcription error rate dropped significantly from 15.58% to 1.11%. This result enabled accurate digitization of complex layouts and specialized terminology that were previously difficult to process.
Additionally, historian Maurice Casey developed ArtificeOCR, a tool connecting olmOCR 2 with Tropy, an open-source photo management application he uses. This allowed him to search thousands of photographs of letters and newspapers related to early 20th-century European radical politics. Casey noted the model performs well with standardized forms in secret police archives and cursive scripts across multiple languages, though he also reported errors when transcribing handwriting in languages less represented in the training data.