# Ai2 reports olmOCR 2 transcription error fell from 15.58% to 1.11% after training on 57 records

_Published Wednesday, October 28, 2026 at 12:22 PM EDT · Science, AI · Latest · Tier 2 — Notable_

![Ai2 reports olmOCR 2 transcription error fell from 15.58% to 1.11% after training on 57 records — Primary](https://www.datocms-assets.com/64837/1793156211-olmocr-featured.png)

Ai2 reports that researchers adapting olmOCR 2 to handwritten pottery records reduced transcription error from 15.58% to 1.11% on 450 test records after training on 57 examples.

The team used LoRA, a method that trains a small set of additional model parameters, to learn the archive's handwriting and abbreviated German terminology. The result comes from CENTURIA's 507-record benchmark.

The model's available weights and code let heritage institutions run and adapt it locally. The team's next goal is processing the roughly 70,000 pottery records in the Carnuntum archive.

## Sources

- [Ai2 Blog](https://allenai.org/blog/olmocr-2-historical-archives)

---
Canonical: https://techandbusiness.org/newswire/HNSK60WuPDg8F97PmlfRmP
Published: 2026-10-28T16:22:44.182Z
Story chronology: 2026-10-28T08:00:00.000Z
Retrieved: 2026-10-28T18:53:40.904Z
Publisher: Tech & Business (techandbusiness.org)