This is artificial Faroese OCR training data created by collection real OCR errors and inserting them into 38 million tokens of non-OCRed text.

The parallel data is set up as a TSV file with the first column (fo_err) being the text with OCR errors, while the second column (fo_corr) is without OCR errors.

This dataset was created by using scripts from https://github.com/atlijas/ocr-post-processing.

Two ByT5 models have been fine-tuned with the data: https://huggingface.co/svanhvit/byt5-ocr-post-processing-faroese-ai-yfirlestur and https://huggingface.co/svanhvit/byt5-ocr-post-processing-faroese

This is a work on progress and more version of the dataset and the models are on the way.