BullingerDB: A Dataset for Handwritten Text Recognition and Writer Retrieval

This dataset provides machine-readable and image-based resources from the Bullinger Digital project, a digital scholarly edition of the correspondence and writings related to the Swiss reformer Heinrich Bullinger (1504–1575). The collection aims to provide a final version of the correspondance ready to use as a benchmark dataset for tasks in the domain of historical document analysis. The handwriting consists of multiple languages - mostly Latin and Old German - but other languages such as Greek also occur. The release contains manuscript page images, line images including the transcription, metadata information (writer, year), and train/test splits for the task of handwritten text recognition and writer retrieval. The statistics for the two tasks proposed - handwritten text recognition and writer retrieval - are shown below. More statistics as well as information about how transcriptions were obtained can be found in our ICDAR2026 paper. Statistics for the task of handwritten text recognition Training Test Frequent Test Non-Frequent Writers 144 144 652 Letters 7,834 939 1,529 Pages 15,619 1,957 3,311 Lines 376,582 48,197 74,443 Statistics for the task of writer retrieval Training Test Writers 164 632 Letters 2,142 8,160 Pages 4,396 16,502 Downloadable files bullingerdb-pages.tar.gz High-resolution manuscript page facsimiles. bullingerdb-lines.tar.gz Cropped line images extracted from manuscript pages. bullingerdb-wr.tar.gzWriter Retrieval page split. The facsimiles are binarized. train.tsvTraining split for HTR. freq_test.tsvFrequent-writer test evaluation split. nonfreq_test.tsvNon-frequent-writer test evaluation split. letters.csvMetadata describing the writer and the year information for every letter. persons.csvMapping from writer ID to the actual person. Information for usage The collection consists of three different hierarchies: Letters, pages and lines. While the facsimile and the line images are sorted according to the letter (the directory 1-1000 contains the letters 1-1000, and each letter has its own directory and lists the pages for this letter), the writer retrieval split is split into a train and test split (WRITER-LETTER-PAGE.png). The meta information (writer and year information) is included in letters.csv. ├── bullingerdb-pages/│ ├── 1-1000/│ │ ├── 1/│ │ │ ├── 0001.png│ │ │ ├── 0002.png│ │ │ └── ...│ │ ├── 2/│ │ └── ...│ ├── 1001-2000/│ └── ... For example, 1-1000/1/0002.png refers to the second page of letter 1. ├── bullingerdb-lines/│ ├── 1-1000/│ │ ├── 1/│ │ │ ├── 0001_1.jpg│ │ │ ├── 0001_2.jpg│ │ │ ├── ... For example, 1/0001_2.jpg refers to the second line of the first page of letter 1. ├── bullingerdb-wr/│ ├── train/│ │ ├── 257_10959_0001.png│ │ ├── 260_11006_0001.png│ │ └── ...│ ├── test/│ │ ├── 6_4016_0001.png│ │ ├── 6_4016_0002.png│ │ └── ... For example, 6_4016_0001.png refers to the first page of letter 4016, written by p6 (see letters.csv). Potential use cases The dataset may be used for handwritten text recognition, writer identification and attribution, writer adaption of computation paleography in general. The benchmarks for HTR and writer identification are listed in the corresponding paper published at ICDAR 2026. Licensing and reuse If you use the dataset, we request you to include a citation to our ICDAR 2026 publication in which the resource is introduced. Additionally, the facsimiles (bullingerdb-pages) may be freely used for scientific, private, non-commercial, and commercial purposes provided the following information is included: the holding archive (Ort der Aufbewahrung) the corresponding archival signature (Signatur) a link to the Creative Commons Attribution-ShareAlike 4.0 International license: Creative Commons BY-SA 4.0 Whenever possible, please also reference the Bullinger Digital project: Bullinger Digital Project - You can also find further information about the Bullinger corpus there.