LaTeX OCR: 200,000 Printed and Synthetic Handwritten Formulas

Holding approximately 200,000 image-text pairs for LaTeX optical character recognition, featuring both printed and synthetic handwritten mathematical formulas. Developed by linxy and updated in December 2024, the dataset aggregates data from Zenodo, CROHME, and custom-built sources to support the LaTeX_OCR project.

Use Cases

• Training image-to-text models to translate printed mathematical notation into LaTeX strings
• Fine-tuning handwriting recognition models using the synthetic_handwrite subset
• Evaluating model performance on complex mathematical syntax using the small test set

Strengths

• ~100,000 printed formula samples
• ~100,000 synthetic handwritten samples for domain adaptation
• AST-based filtering to ensure all LaTeX strings are renderable

Limitations

• Handwritten data is synthetic (font-based) rather than human-sourced
• Small test set of only 110 samples
• Specific column names are not documented in the metadata

Provenance

Source

Zenodo (record 56198), CROHME (isical.ac.in), and LinXueyuanStdio/Data-for-LaTeX_OCR

Collection Method

Aggregated from existing repositories and synthetically generated using handwriting fonts

Freshness

Last updated December 2024