The OCR text (and ABBYY) for about 15 millions are already available today, e.g. c.f. https://archive.org/advancedsearch.php?q=format:abbyy&fl=identifier,format

The OCR text (and ABBYY) for about 15 millions are already available today, e.g. c.f. https://archive.org/advancedsearch.php?q=format:abbyy&fl=identifier,format&rows=25&output=json&sort=downloads+desc https://archive.org/advancedsearch.php?q=format:abbyy&fl=ide... The ABBYY files are raw output of OCR, but most "items" on the Archive also include the extracted plain text. It's noisy but generally quite readable. A few heuristics would pass pretty clean copy to GPT-2.