kbd-ru-opus

Kabardian-Russian Translation Model (kbd-ru-opus)

Fine-tuned MarianMT model for Kabardian (East Circassian) to Russian translation.

• Developed by: Eduard Emkuzhev
• Model type: Neural Machine Translation (Marian Transformer)
• Language pair: Kabardian (kbd) → Russian (ru)
• License: CC BY-NC 4.0
• Base model: Helsinki-NLP/opus-mt-en-ru (Apache 2.0)
• Training data: adiga-ai/circassian-parallel-corpus (CC BY 4.0)

Model Description

This model translates from Kabardian to Russian. Kabardian is an endangered Northwest Caucasian language with approximately 500,000 speakers. It features complex polysynthetic morphology, 50+ consonants, and ergative-absolutive alignment.

Intended Use

Primary uses:

• Language documentation and digitization
• Educational content translation
• Cultural heritage preservation
• Low-resource NMT research
• Supporting Kabardian speakers in accessing Russian content

Limitations:

• Non-commercial use only (CC BY-NC 4.0)
• Best performance on everyday language
• May struggle with modern technical terms not in training data
• Requires proper handling of Kabardian-specific character Ӏ (palochka)

Training Data

Dataset: adiga-ai/circassian-parallel-corpus

• Subset: kbd_ru (Kabardian → Russian)
• Total training examples: ~120K parallel sentence pairs
• Dataset license: CC BY 4.0
• Dataset author: Anzor Qunash (adiga.ai)
• Content: Dictionary entries, folklore texts, proverbs, everyday expressions

Training Procedure

Base Model

• Architecture: Marian Transformer (transformer-align)
• Base: Helsinki-NLP/opus-mt-en-ru (English-Russian translation)
• Transfer learning: Adapted from English-Russian to Kabardian-Russian

Hyperparameters

base_model: Helsinki-NLP/opus-mt-en-ru
training_examples: 200,000
epochs: 7
batch_size: 32
learning_rate: 3e-4
optimizer: AdamW
max_sequence_length: 128
warmup_steps: 500
weight_decay: 0.01
framework: transformers 4.36.0

Special Preprocessing

The model uses a special character mapping for training:

• Kabardian Ӏ (palochka) → I (Latin I) during training
• I → Ӏ restored during inference

This ensures better tokenization compatibility with the MarianMT tokenizer.

Performance

Benchmark Results

Tested on 1,000 examples from adiga-ai/circassian-parallel-corpus:

Test Configuration:

• Test size: 1,000 examples
• Sampling: Every 50th sentence from corpus
• Generation: beam_search (num_beams=4)
• Device: Apple M-series (MPS)
• Seed: 42 (reproducible)

Translation Examples

Note: The model successfully handles complex Kabardian morphology and preserves meaning in Russian translations.

How to Use

Installation

pip install transformers torch sentencepiece

Basic Usage

from transformers import MarianMTModel, MarianTokenizer

# Load model and tokenizer
model_name = "kubataba/kbd-ru-opus"
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

# Translation function
def translate_kbd_to_ru(text):
    # Preprocess: Ӏ → I for tokenization
    processed_text = text.replace('Ӏ', 'I').replace('ӏ', 'I')
    
    inputs = tokenizer(processed_text, return_tensors="pt", padding=True)
    outputs = model.generate(**inputs, max_length=128, num_beams=4)
    translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return translation

# Example
kabardian_text = "Сэлам!"
russian_text = translate_kbd_to_ru(kabardian_text)
print(f"KBD: {kabardian_text}")
print(f"RU: {russian_text}")

Batch Translation

texts = [
    "Уи пщэдджыжь фӀыуэ!",
    "Сыт укъэпсэлъар?",
    "Ди лъэпкъым и бзэр"
]

# Preprocess all texts
processed_texts = [t.replace('Ӏ', 'I').replace('ӏ', 'I') for t in texts]

inputs = tokenizer(processed_texts, return_tensors="pt", padding=True, truncation=True)
outputs = model.generate(**inputs, max_length=128, num_beams=4)
translations = [tokenizer.decode(out, skip_special_tokens=True) for out in outputs]

for src, tgt in zip(texts, translations):
    print(f"KBD: {src} → RU: {tgt}")

Limitations and Bias

• Complex morphology: Kabardian's polysynthetic structure may result in one-to-many mappings to Russian
• Character handling: Requires preprocessing of palochka (Ӏ → I) before tokenization
• Domain adaptation: Best performance on text types similar to training data
• Technical vocabulary: May struggle with modern technical/specialized terms
• Endangered language: Limited real-world validation data due to language endangerment
• Dialectal variation: Trained on literary Kabardian; dialectal forms may produce unexpected results

Ethical Considerations

This model contributes to digital language preservation for Kabardian, an endangered language.

Important considerations:

• Machine translation should complement, not replace, human translators
• Cultural sensitivity is essential when working with indigenous languages
• The model may not capture all nuances of Kabardian language and culture
• Translations should be reviewed by native speakers for critical applications
• Supporting Kabardian language education and preservation is crucial

About Kabardian Language

Kabardian (Adyghe-Kabardian, East Circassian) is a Northwest Caucasian language spoken by approximately 500,000 people in:

• Kabardino-Balkaria (Russia)
• Karachay-Cherkessia (Russia)
• Turkey (diaspora communities)
• Middle East (diaspora communities)

Linguistic features:

• Phonology: 50+ consonant phonemes (one of the world's largest inventories)
• Morphology: Polysynthetic - complex word formation
• Syntax: Ergative-absolutive alignment
• Writing: Cyrillic script + palochka (Ӏ) for glottal stop
• Status: Endangered (UNESCO classification)

License and Attribution

This Model

• License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
• Author: Eduard Emkuzhev
• Year: 2025

Base Model

• Model: Helsinki-NLP/opus-mt-en-ru
• License: Apache 2.0
• Authors: Language Technology Research Group at the University of Helsinki
• Link: https://huggingface.co/Helsinki-NLP/opus-mt-en-ru

Training Dataset

• Dataset: Circassian-Russian Parallel Corpus v1.0
• License: CC BY 4.0
• Author: Anzor Qunash (adiga.ai)
• Link: https://huggingface.co/datasets/adiga-ai/circassian-parallel-corpus

Citation

If you use this model in your research, please cite:

@misc{emkuzhev2025kbdru,
  author = {Eduard Emkuzhev},
  title = {Kabardian-Russian Neural Machine Translation Model},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/kubataba/kbd-ru-opus}}
}

Please also cite the base model and dataset:

@misc{helsinki-nlp-opus-en-ru,
  author = {Language Technology Research Group at the University of Helsinki},
  title = {OPUS-MT English-Russian Translation Model},
  year = {2020},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/Helsinki-NLP/opus-mt-en-ru}}
}

@dataset{qunash2025circassian,
  author = {Anzor Qunash},
  title = {Circassian-Russian Parallel Text Corpus v1.0},
  year = {2025},
  publisher = {adiga.ai},
  url = {https://huggingface.co/datasets/adiga-ai/circassian-parallel-corpus}
}

Acknowledgments

• Helsinki-NLP team for the excellent OPUS-MT base models
• Anzor Qunash (adiga.ai) for creating and publishing the Circassian-Russian Parallel Corpus
• Kabardian language community for preserving and promoting their language
• All contributors to Circassian language digitization efforts

Related Models

• Reverse direction: kubataba/ru-kbd-opus - Russian to Kabardian
• Base model: Helsinki-NLP/opus-mt-en-ru
• Dataset: adiga-ai/circassian-parallel-corpus

Technical Details

• Framework: PyTorch + Transformers
• Model size: ~300MB
• Vocabulary size: ~62,5K tokens
• Parameters: ~74M
• Inference: CPU and GPU compatible
• Optimal device: GPU or Apple Silicon (MPS)

Contact

• Author: Eduard Emkuzhev
• Email: [email protected]
• GitHub: https://github.com/kubataba
• Issues: Report issues

For commercial licensing inquiries, please contact via email.

---

Model Card Authors: Eduard Emkuzhev

Last Updated: December 2025

Version: 1.0.1