LightOn

Legacy OCR Texts Make Language Models Learn at Only 30 Percent Efficiency – the FineBooks Project Aims to Change That

Hugging Face and EleutherAI's FineBooks project tested 14 open OCR models on more than 2,000 historical book pages. The results show that smaller models often outperform larger competitors, with the top model achieving over 97% character accuracy at less than $2 per thousand pages. According to the authors, the quality is sufficient for AI training but not yet for scholarly purposes.

Hugging Face EleutherAI PaddleOCR (Baidu) LightOn DeepSeek