TextMuSS-10M

TextMuSS-10M is a large-scale synthetic multilingual scene text dataset introduced in:

All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts
Xingsong Ye, Yongkun Du, Jiaxin Zhang, Zhixian Li, Chong Sun, Chen Li, Jing Lyu, Lianwen Jin, Zhineng Chen
arXiv:2609.24058 (2026)

• 📄 Paper: https://arxiv.org/abs/2609.24058
• 🤗 Hugging Face Paper: https://huggingface.co/papers/2609.24058
• 💻 Code: https://github.com/YesianRohn/ScriptMoE
• 🚀 Demo: https://huggingface.co/spaces/Yesianrohn/MultilingualOCR-Demo
• 📊 Benchmark: https://huggingface.co/datasets/Yesianrohn/TextMuSS-Bench

Dataset Summary

TextMuSS-10M was created to provide large-scale and balanced multilingual supervision for scene text recognition.

Real-world scene text datasets are heavily imbalanced across languages and scripts, with many languages having little or no high-quality training data. TextMuSS-10M addresses this limitation through large-scale synthetic scene text generation.

The dataset covers:

• 10 writing scripts
• 229 languages
• approximately 1 million synthetic samples per script
• approximately 10 million samples in total

The dataset is used as a major source of multilingual supervision for ScriptMoE.

Motivation

Multilingual scene text recognition requires models to handle visually and linguistically diverse writing systems.

However, real-world training resources are concentrated around a relatively small number of high-resource languages. This creates a long-tail problem for multilingual OCR.

TextMuSS-10M provides synthetic supervision across a much broader set of languages and scripts, allowing a single recognizer to learn multilingual scene text representations.

Relationship to ScriptMoE

TextMuSS-10M is the main synthetic multilingual training resource introduced alongside ScriptMoE.

ScriptMoE uses a shared visual encoder and a script-aware sparse Mixture-of-Experts decoder. The dataset provides multilingual supervision for the different script-specialized experts while also supporting cross-script knowledge learning through the shared expert.

Dataset Structure

The exact file organization and metadata format are available in the dataset repository.

The dataset is hosted in Hugging Face Xet storage due to its large size.

Intended Use

TextMuSS-10M is intended for:

• multilingual scene text recognition training
• OCR research
• script-aware representation learning
• multilingual vision-language research
• robustness studies for low-resource scripts and languages
• training and evaluating multilingual OCR systems

Limitations

TextMuSS-10M is a synthetic dataset.

Synthetic scene text may not fully reproduce the visual diversity, image degradation, typography, layout, and environmental conditions found in real-world photographs.

Therefore, models trained on TextMuSS-10M should ideally be evaluated on real-world multilingual benchmarks.

For real-image evaluation, see TextMuSS-Bench.

Dataset Size

The Hugging Face repository currently contains approximately 83.2 GB of data.

License

This dataset is released under the Apache-2.0 license.

Please also review the licenses and terms of any fonts, language resources, or third-party components used in downstream processing or redistribution.

Citation

If you use TextMuSS-10M, please cite the accompanying paper:

@article{ye2026scriptmoe,
  title   = {All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts},
  author  = {Ye, Xingsong and Du, Yongkun and Zhang, Jiaxin and Li, Zhixian and Sun, Chong and Li, Chen and Lyu, Jing and Jin, Lianwen and Chen, Zhineng},
  journal = {arXiv preprint arXiv:2609.24058},
  year    = {2026}
}

Paper: https://arxiv.org/abs/2609.24058

Hugging Face Paper: https://huggingface.co/papers/2609.24058