Skip to content

Dataset · lightonai

Large full-page OCR dataset

lightonai/LightOnOCR-mix-0126

10m-100m OCR examples: full page transcriptions in Markdown with LaTeX and HTML tables for end-to-end OCR training.

What's inside

The dataset was gathered via distillation: a strong vision-language model generates naturally ordered transcriptions of rendered document pages. Inside are full page transcriptions in Markdown, including LaTeX formulas and HTML tables. Languages: en, fr, de, es, it, ja. Size ranges from 10m to 100m examples.

What you can do with it

  • Training end-to-end OCR models on full pages
  • Fine-tuning vision-language models for document understanding
  • Checking transcription quality of Markdown with formulas and tables
Tasks
Image → text
Languages
English, French, German, Spanish, Italian, Japanese, Russian, Polish and 4 more
Size
10M–100M rows
Downloads this month
1.4K
Likes
115

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("lightonai/LightOnOCR-mix-0126")