Glyphwise recognition models are trained on the sources below. No training data is included in the distribution; this file records provenance and the attributions the data licences ask for.
Text and documents
- Project Gutenberg texts (public domain in the United States): sentences for rendered training lines and the language model. Project Gutenberg trademarks and boilerplate are stripped and not used.
- US federal government documents (public domain under 17 U.S.C. 105): IRS forms, instructions and publications; the Federal Register; NIST and GPO publications; Supreme Court opinions; public laws, House bills, the Code of Federal Regulations and the US Code. Their text layers label line images rendered from the PDFs (
tools/pd_sources.txt,tools/pd2_sources.txt).
Handwriting (Creative Commons Attribution 4.0, https://creativecommons.org/licenses/by/4.0/)
- GNHK — GoodNotes Handwriting Kollection. GoodNotes, A. W. C. Lee, J. Chung, M. Lee: "GNHK: A Dataset for English Handwriting in the Wild", ICDAR 2021.
- Transcriptorium / Bentham — line images and transcripts from the Bentham Papers (UCL Transcribe Bentham), as released by the tranScriptorium/READ projects.
- READ datasets (Zenodo) — "ICDAR 2017 Competition on Handwritten Text Recognition on the READ Dataset" (Train-A) and the READ Bozen/Bolzano dataset.
Line images from these datasets were cropped, normalized and augmented for training; the datasets themselves are not redistributed.
Fonts
Synthetic training lines are rendered from font files. The model shipped with the product (line_clean1c, and line_clean1b and line_clean1 before it) was rendered only from the Google Fonts collection (its ofl/, apache/ and ufl/ folders: fonts under the SIL Open Font License 1.1, the Apache License 2.0 and the Ubuntu Font Licence 1.0), which permit use of the fonts for any purpose. Development models have also been trained on other fonts (for example those installed with Windows) and must not be distributed.
Font licences differ in whether rendering a font to train a model is permitted; review the licences of the fonts used before distributing any other model commercially.