Latex Ocr: 10,000+ Mathematical Image-to-LaTeX Pairs

Between 10,000 and 100,000 image-text pairs comprise this 1% sample of the LaTeX_OCR collection for mathematical formula recognition. Released by unsloth in late 2024, the data provides paired visual representations and LaTeX source code formatted for Polars and Pandas.

Use Cases

• Training OCR models to map mathematical images to LaTeX strings
• Evaluating vision-to-text accuracy for complex mathematical symbols
• Fine-tuning multimodal models on structured mathematical notation

Strengths

• Apache 2.0 permissive license
• 10,000 to 100,000 record count
• Parquet format for high-performance data loading

Limitations

• 1% sampling rate significantly reduces the diversity found in the parent dataset
• Lack of explicit column documentation in the source metadata

Provenance

Source

https://huggingface.co/datasets/linxy/LaTeX_OCR

Collection Method

Sampled from an existing dataset

Freshness

Last updated November 2024.