This dataset is provided solely for the purpose of participating in the R.O.A.D. Barbados Historic Handwriting Challenge hosted on Zindi. Use of this dataset outside of this scope is strictly prohibited.

You may not copy, distribute, transmit, publish, or use this dataset for any other research, commercial, educational, or public purpose. This includes, but is not limited to, uploading to public repositories or using for other competitions.

Violation of this license may result in disqualification and potential legal action.

The dataset consists of 6K cropped images of lines of text, extracted from scanned historical records in Barbados’ archive. These are real handwritten words from the 18th and 19th century: partial scans of deeds, wills, inventories, and other legal or estate records.

Each image contains one or more handwritten words. The images are split into training and test sets, and the goal is to train a model that can correctly transcribe these handwritten words into digital text.

Note: These aren’t modern handwriting samples. You should expect old penmanship styles, variable spacing, as well as the effects of time on paper such as faded ink, damage and other imperfections. Your model will need to be robust to real-world imperfections you'd encounter when working with historical documents.