GlotOCR-bench

GlotOCR-bench is a dataset of 16375 images covering 158 writing systems (+2000 languages), designed to evaluate the fundamental OCR capabilities required to support diverse writing systems and languages.

Quick links:

• 📃 Paper
• 🛠️ Code
• 📈 Results

License

This dataset is released under the GlotOCR Open Evaluation License v1.0 (see LICENSE file for full terms).

• The GlotOCR-bench metadata is licensed under CC0-1.0.
• The texts used to generate the images come from the GlotLID repository. Each text is accompanied by a source tag and remains licensed under its own terms. You can find the list of sources (match the key) and their licenses here: https://github.com/cisnlp/GlotLID/blob/main/sources.md
• The licenses of all selected source texts permit their use for evaluation.
• Allowed: Evaluation and benchmarking & Internal research
• Not allowed: Training or fine-tuning models (commercial or non-commercial) & Creating training datasets from this data

This dataset is intended only for evaluation purposes.

Code

The image generation code is released under the Apache-2.0 license and may be used to create training data, provided that different seed texts are used from those in this dataset.

Contact

For specific use cases or licensing questions: amir@cis.lmu.de

Citation

@misc{kargaran2026glotocrbench,
      title={GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts}, 
      author={Amir Hossein Kargaran and Nafiseh Nikeghbal and Jana Diesner and François Yvon and Hinrich Schütze},
      year={2026},
      eprint={2604.12978},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.12978}, 
}