This model is a Kraken text recognition model trained to recognize Syriac manuscript text in three scripts: Serto, Estrangela and Eastern. This model is the result of a Syriac Transcribathon held at Princeton University in 2023 and organized by researchers from Université Paris Sciences & Lettres, Princeton University, and Beth Mardutho. Syriac manuscripts were sourced from the following repositories (in decreasing order of manuscript usage): - Library of Congress - Vatican Library - National Library of France - Harvard Library - Internet Archive - Princeton University Library - British Library Model Metrics - 90,991 characters - 3,466 lines - 100 images (both full color digital photos and greyscale microfilms) - mostly Biblical texts. - 38 manuscripts whose dates ranged between the 6th and 20th centuries. Characters permitted in the HTR ground truth were limited to the Syriac consonants (excluding the extra letters for Garshuni or Syro-Persian), space, a very limited selection of interpunctuation and an even more limited selection of diacritics, and no vowels. Spaces are provided in the transcriptions even when spacing was not consistently present in the source manuscript. Besides regularizing the spacing, no further regularizations were made to the transcriptions. During model training, it achieved an accuracy score of 97.4% on the test data. For both tasks combined (layout segmentation, text recognition), participants in the Transcribathon annotated a total of 1,021 images across 161 manuscripts that Funding sources - ERC, MiDRASH, Project No. 101071829 - Princeton Department of Near Eastern Studies - Princeton Center for Digital Humanities - Princeton Manuscript, Rare Book & Archival Studies - Biblissima+, Cluster 3, Project ANR-21-ESRE-0005. Contributors Data Collectors Christopher Mrani, Marianne Farraj, Marie Hanna, Gabriel Mrani, Maria Thomas, Gabriel Rabo, Saranya Chandran, and Helene Rey. Researchers George Kiraz, Christine Roughan, and Daniel Stökl Ben Ezra.