Hugging Face's FineBooks: The Best OCR for Historical Books Costs Under $2 per Thousand Pages

Hugging Face, together with EleutherAI, has published the results of its FineBooks project: the team tested 14 open OCR models on more than 2,000 pages of historical books. The best solution it found, the dots.mocr model, reaches 97.6% character-level accuracy at a cost of under two dollars per thousand pages.

Why historical books need re-digitizing

Language models train on huge text corpora, a significant share of which consists of digitized books from public archives. The problem is that most of these books were scanned in the 1990s and 2000s with primitive OCR software. The result is thousands of systematic errors: swapped letters, broken words, lost lines, unreadable abbreviations.

When this "dirty" OCR text ends up in a training set, the model learns from distorted patterns. That lowers generation quality, especially in domains where historical texts make up a large share: law, medicine, classic literature, the history of science.

FineBooks is an attempt by Hugging Face and EleutherAI to tackle this problem systematically. The team deliberately chose historical book pages as the hardest OCR target: nonstandard old typefaces, yellowed paper, uneven ink, and printing defects typical of editions from past centuries. Before FineBooks, there was no systematic comparison of open OCR solutions on historical material specifically.

Which model came out on top

The test covered 14 open OCR systems on a sample of more than 2,000 pages from historical editions. The winner, dots.mocr, scored 97.6% character-level accuracy. That means fewer than 3 wrong characters per 100, the threshold at which text becomes usable for building training datasets for large language models.

Key figures from FineBooks:

• 14 open OCR models compared under a single methodology
• More than 2,000 historical book pages in the test sample
• Leader: dots.mocr, with 97.6% character-level accuracy
• Processing cost: under $2 per thousand pages

The economics matter. At under $2 per thousand pages, reprocessing a corpus of a million pages costs less than $2000, a realistic budget for non-profit AI research organizations building open datasets.

Where accuracy still falls short

The FineBooks authors openly acknowledge the limitations. A 97.6% score is sufficient for preparing AI training data, but not for academic transcription. Scholarly transcription requires near-perfect reproduction of the original: historians and textual scholars work with texts in which every comma carries meaning, and their acceptable error rate is an order of magnitude lower than for AI datasets.

"The results are good enough for AI training data, but do not yet reach the level required for scholarly transcription," according to the

FineBooks project materials, published jointly by Hugging Face and EleutherAI.

This is an important distinction: the study states the limits of dots.mocr's applicability plainly and does not present it as a universal solution. dots.mocr meets the needs of the AI industry, but the academic community is still waiting for higher accuracy.

What this means

FineBooks gives the industry a concrete working benchmark: dots.mocr is a cost-effective tool for large-scale re-digitization of historical archives for LLM training. According to Hugging Face and EleutherAI, the problem of "dirty" OCR text in training corpora is real and solvable today, without waiting for expensive commercial solutions. FineBooks opens the way for millions of pages of historical books to finally become a full part of the training corpora of open language models.