k
keynews.ai
keynews.ai
FrontierThe story, in brief

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

97.6% accuracy, $2 per 1K pages—Hugging Face and EleutherAI solve the OCR bottleneck starving LLM training on historical texts.

Reporting by The Decoder 1 min brief
Illustration of a transparent lens revealing connected networks across layers of paper.
Exploring the next frontier of AI research.AI illustration by KeyNews
The KeyNews take

Why it matters

Training-data quality is a hard constraint on frontier models. FineBooks benchmarks OCR at scale and proves cost-effective solutions exist—practitioners sourcing historical text corpora now have a playbook.

The key facts

12 to know

  1. FineBooks tested 14 open-source OCR models

  2. 2,000+ historical book pages evaluated

  3. Top performer: dots.mocr at 97.6% character accuracy

  4. Cost: under $2 per 1,000 pages

  5. Suitable for AI training (not yet for scholarly transcription)

  6. Hugging Face + EleutherAI collaboration

  7. FineBooks tested 14 open-source OCR models on 2,000+ historical book pages

  8. Top model (dots.mocr) achieves 97.6% character accuracy

  9. Cost: under $2 per thousand pages

  10. Accuracy sufficient for AI training; not yet for scholarly transcription

  11. Partnership: Hugging Face and EleutherAI

  12. Focus: removing OCR degradation as a training-data constraint

Go to the source

The Decoderthe-decoder.com

Publisher excerpt: The FineBooks project from Hugging Face and EleutherAI tested 14 open-source OCR models on more than 2,000 historical book pages. The top model, dots.mocr, hits 97.6 percent character accuracy at under two dollars per thousand pages. That's good enough for AI training data, but not yet for…
Read original report
Back to today's editionMore frontier news

The wider picture

View all
Illustration of a transparent lens revealing connected networks across layers of paper.
AI illustration by KeyNews
Frontier01

Alibaba Unveils Zhenwu V900 — and Plans Qwen Models With Up to 10 Trillion Parameters

Alibaba is advancing on two fronts simultaneously: announcing a custom AI accelerator (Zhenwu V900) and committing to massive model scale (10T parameters for future Qwen releases). For practitioners, this matters as a credible third-party capability play outside the US-China licensing squeeze; for enthusiasts, it's a significant lab-race signal about training compute and parameter scaling as competitive levers.

TechRepublic
Illustration of a transparent lens revealing connected networks across layers of paper.
AI illustration by KeyNews
Frontier02

Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning

A new capability frontier: models that reason directly in speech without transcription bottlenecks. This changes how we think about multimodal reasoning and what's possible with open-weight releases at scale.

MarkTechPost
Illustration of a transparent lens revealing connected networks across layers of paper.
AI illustration by KeyNews
Frontier03

Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war

Frontier labs are shipping upgraded reasoning and multimodal models in rapid succession, signaling acceleration in the capability race. Simultaneous price cuts reshape AI economics for practitioners.

Simon Willison