MixTeX Synthetic Pretraining Dataset

This record contains the synthetic pretraining dataset used for MixTeX, a data-efficient LaTeX OCR system introduced in our ICIP 2026 paper, "MixTeX: Data-Efficient LaTeX OCR via Synthetic Pretraining and Limited Fine-Tuning".

The dataset is generated without using real LaTeX document sources. It randomly pairs grammatical Wikipedia text with grammatical LaTeX formulas to create syntactically valid bilingual LaTeX OCR training samples. The dataset includes English and Chinese synthetic pretraining data. The generated samples are intended for training image-to-LaTeX OCR models and for supporting research on synthetic data generation, multilingual document understanding, and low-resource LaTeX OCR.

Official Repository: https://github.com/yuhanxu01/MixTeX

Citation: If you find this dataset, our code, or our paper useful in your research, please consider citing:

@inproceedings{xu2026mixtex, title={MixTeX: Data-Efficient LaTeX OCR via Synthetic Pretraining and Limited Fine-Tuning}, author={Xu, Yuhan and Zhao, Yijun and Luo, Renqing and Weiss, Gary M.}, booktitle={2026 IEEE International Conference on Image Processing (ICIP)}, year={2026}, organization={IEEE} }

@misc{mixtex, title = {MixTeX code and data repository}, author = {Xu, Yuhan and Zhao, Yijun and Luo, Renqing and Weiss, Gary M.}, year = {2025}, howpublished = {\url{https://github.com/yuhanxu01/MixTeX}} }