LightOnOCR-bbox-mix-0126: Large-Scale OCR Training Data with Layout

LightOnOCR-bbox-mix-0126 is a large-scale OCR training dataset built via distillation from a strong vision-language model. The dataset provides naturally ordered full-page transcriptions in Markdown format, including LaTeX math spans and HTML tables, designed for end-to-end document understanding models. It was created by lightonai and last updated on Hugging Face in January 2026.

Use Cases

• Training OCR models to output clean, human-readable text based on the described full-page transcriptions.
• Developing document-understanding models that preserve layout and structure based on the inclusion of Markdown, LaTeX, and HTML elements.
• Fine-tuning vision-language models for document parsing tasks based on the distillation method used to create the dataset.

Strengths

• Dataset is described as large-scale, suggesting substantial volume for training.
• Includes layout information and complex document elements like LaTeX math and HTML tables.
• Created via distillation from a strong vision-language model, which may indicate high-quality synthetic labels.

Limitations

• Row count, file formats, and column-level documentation are unknown, limiting suitability assessment.
• License information is not provided, which is critical for commercial use.
• Data may reflect biases inherent to the source documents and the distillation model used.

Provenance

Source

lightonai on Hugging Face.

Collection Method

Built via distillation: a vision-language model was prompted to produce transcriptions from rendered document pages.

Freshness

Last updated 2026-01-21 10:51:53.