ScaDS.AI German Line- and Word-Level Handwriting Dataset

This dataset contains 775 full-page images, 5.843 line images and 37.406 word images with German handwritings and corresponding line- and word-level ground truth text files. There are no ground truth text files for the full-page images. If you need full-page images with ground truth, check out our dataset with the name „ScaDS.AI German full-page Handwriting Dataset“ on Zenodo (includes 77 pages). Structure - /ground_truth/xml: 775 ground truth xml files - /ground_truth/csv/line_annotations.csv: ground truth texts - /ground_truth/csv/word_annotations.csv: ground truth texts - /images/pages: 775 full-page images with German handwriting - /images/lines: 5.843 line images with German handwriting - /images/words: 37.406 word images with German handwriting - /metadata/Uni_Leipzig_01-export-251205.csv (further descriptions, e.g. corresponding wikipedia urls) Description We offer ground truth files both in XML format and in CSV format with annotated words and lines which cover about 90% of the pages (relating on the content in Uni_Leipzig_01-export-251205.csv). Annotating partially unreadable hundred of manuscripts (775) is hard work and takes a lot of time. We provide about 43.249 annotated lines and words from excerpts of Wikipedia articles with ground truth text. Handwriting OCR in German is underserved. The quality is not perfect. We did our best, but still there are mistakes. A third party review of the dataset (only line labels) produced the following results (automatically checked by a vision model, which took place in August 2026): 87% were exact character-for-character matches. 11% contained minor discrepancies, mostly involving punctuation or silently corrected spelling mistakes made by the original writers. 2% were clearly mislabeled: the label either did not match the corresponding line or omitted written words. These results were provided by a third party (using a vision model). I have not independently verified them and therefore cannot guarantee their accuracy. Version * 1.0 created on January 19, 2026 (10.5281/zenodo.18301532) Authors * Thomas Burghardt (0009-0003-7874-9134) * Ahmad Alzin (0009-0006-4791-5939) Licence Creative Commons Attribution 4.0 International Use cases - Handwritten Text Recognition (HTR) - OCR training - Computer vision / handwriting analysis Acknowledgements We acknowledge the financial support by the Federal Ministry of Education and Research of Germany and by Sächsische Staatsministerium für Wissenschaft, Kultur und Tourismus in the programme Center of Excellence for AI-research „Center for Scalable Data Analytics and Artificial Intelligence Dresden/Leipzig“, project identification number: ScaDS.AI