CrossLing-OCR-Mini

๐Ÿš€ CrossLing-OCR-Mini is a lightweight OCR model designed for low-resource multilingual languages.


1. Model Overview

Despite its compact size (~580MB), the model demonstrates strong recognition performance across 11 languages, while remaining deployable on consumer-grade GPUs.

Key Features

  • Multilingual OCR with structure-aware text recognition
  • Specialized optimization for low-resource and complex scripts
  • Lightweight (~580MB) and efficient inference

Supported Languages

  • High-resource languages: Chinese, English
  • Low-resource languages (specially optimized):
    Tibetan, Mongolian, Kazakh, Kyrgyz, Zhuang, etc

Experimental results indicate that CrossLing-OCR-Mini outperforms or matches mainstream OCR systems on multiple low-resource languages.


2. Usage / Inference

CrossLing-OCR-Mini can be directly used with the ๐Ÿค— Transformers library.
The following example demonstrates single-image OCR inference for plain text recognition.

Requirements

  • Python โ‰ฅ 3.8
  • transformers (latest version recommended)
  • CUDA-enabled GPU (recommended for optimal performance)
pip install -U transformers accelerate

Simple OCR Inference Example

from transformers import AutoModel, AutoTokenizer

# Hugging Face model id
model_id = "NCUTNLP/CrossLing-OCR-Mini"
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True
)
model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    low_cpu_mem_usage=True,
    device_map="cuda",
    use_safetensors=True,
    pad_token_id=tokenizer.eos_token_id
)
model = model.eval().cuda()
# Input image
image_file = "test.png"
# Perform plain text OCR
result = model.chat(
    tokenizer,
    image_file,
    ocr_type="ocr"
)
print("Predicted OCR result:\n")
print(result)

Notes

  • ocr_type="ocr" enables plain text OCR mode
  • The model automatically handles multilingual text recognition
  • For best results, input images should be clear and upright
  • Consumer-grade GPUs (e.g., RTX 3060 / 3090) are sufficient for inference

3. Performance Notes & Limitations

While CrossLing-OCR-Mini achieves strong overall performance, several limitations remain:

  • OCR accuracy on Mongolian and Uyghur still has room for improvement
  • Performance may degrade on extremely noisy, handwritten, or out-of-distribution inputs

These challenges will be addressed in future versions of the model.


4. Model Variants

Version Intended Use Availability
CrossLing-OCR-Mini Research and academic purposes only โœ… Open-sourced
CrossLing-OCR-Pro-Preview Commercial / production purposes ๐Ÿ”’ Contact required

The performance differences between the Mini and Pro-Preview versions are illustrated below.

Mini\_Pro-Preview


5. Prohibited Use & Disclaimer

This model must not be used for:

  • Any illegal or unlawful activities
  • Applications that violate applicable laws or regulations
  • Surveillance or profiling that infringes on individual rights
  • Discriminatory or harmful automated decision-making in sensitive contexts

Disclaimer:

  • Any misuse of this model is solely the responsibility of the user
  • The authors and maintainers do not endorse and are not liable for any consequences arising from improper or malicious use
  • Outputs generated by this model do not represent the views or positions of the authors

6. Ethical Considerations & Bias

CrossLing-OCR-Mini is developed to support research on low-resource and underrepresented languages. However, like all OCR systems, the model may reflect biases present in its training data, including:

  • Uneven performance across languages and scripts
  • Sensitivity to document quality, typography, and layout variations
  • Reduced robustness on degraded, historical, or low-resolution documents

Users are encouraged to:

  • Carefully evaluate outputs before downstream use
  • Avoid deploying the model in high-risk or sensitive decision-making scenarios

7. License

This model is released for research purposes only. Commercial use is not permitted without explicit authorization.

For commercial licensing or extended usage, please contact the authors.


8. Contact

For questions, collaboration, or commercial inquiries:

๐Ÿ“ง [email protected]


Downloads last month
74
Safetensors
Model size
0.6B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support