Panel 2 · Specialised Models, Scalability, and the Methodology of HTR

Navigating bilingual Pardon Letters: custom training vs. specialised supermodels in the PARDONS-Project

A custom model at 30.13% CER wasn't good enough — but a hybrid workflow using the 'Dutch Dean' supermodel changed everything. This talk traces the PARDONS project's pivot from custom training to leveraging specialised supermodels for 16th- and 17th-century bilingual (Dutch-French-Latin) pardon letters from the Belgian State Archives, and shares honest reflections on when to build your own model, and when not to.

Part of

Panel 2 · Specialised Models, Scalability, and the Methodology of HTR