Old OCR text cripples language model training, and FineBooks wants to fix that at scale

FineBooks evaluates 14 open-source OCR models to fix noisy historical text for AI training. The top model, dots.mocr, achieves 97.6% character accuracy at a cost of under two dollars per thousand pages. This research addresses a major data quality bottleneck by identifying reliable pipelines for converting historical books into clean training corpora.

Opening Kapyn…