---
license: mit
datasets:
- skypro1111/ubertext-2-news-verbalized
language:
- uk
base_model:
- skypro1111/m2m100-ukr-verbalization
pipeline_tag: text2text-generation
library_name: transformers
widget:
- text: >-
    Очікувалось, що цей застосунок буде запущено о 11 ранку 22.08.2025, але
    розробники затягнули святкування і запуск був відкладений на 2 тижні.
---

# m2m100-ukr-verbalization-ct2

This is the CTranslate2-optimized version of [skypro1111/m2m100-ukr-verbalization](https://huggingface.co/skypro1111/m2m100-ukr-verbalization).

## About CTranslate2 Optimization

CTranslate2 is an optimized inference engine for Transformer models that provides much faster inference than standard implementations. This repository contains the m2m100-418M model for Ukrainian verbalization, optimized for production deployment.

Key advantages:
- **Faster inference**: 3-5x speed improvement over the standard Transformers implementation
- **Lower memory footprint**: Uses efficient int8/float16 quantization
- **Better hardware utilization**: Optimized for both CPU and GPU
- **Efficient batching**: Improved batch processing for production workloads

## Usage

```python
import os
import ctranslate2
from transformers import M2M100Tokenizer

def process_sentence(translator, tokenizer, sentence: str):
    # Tokenize input
    source = tokenizer.convert_ids_to_tokens(tokenizer.encode(sentence))
    target_prefix = [tokenizer.lang_code_to_token["uk"]]
    
    # Run inference
    results = translator.translate_batch(
        [source],
        target_prefix=[target_prefix],
        beam_size=1,
        num_hypotheses=1,
        use_vmap=True,
    )
    
    # Get target tokens and decode
    target = results[0].hypotheses[0][1:]  # Remove language token
    return tokenizer.decode(tokenizer.convert_tokens_to_ids(target))

# Initialize translator with optimizations
translator = ctranslate2.Translator(
    "path/to/m2m100-ukr-verbalization-ct2",
    device="cuda",
    compute_type="int8_float16",  # Options: int8, int8_float16, float16, float32
    intra_threads=16,
)

# Load tokenizer from the original model
tokenizer = M2M100Tokenizer.from_pretrained("skypro1111/m2m100-ukr-verbalization")
tokenizer.src_lang = "uk"

# Example
text = "Моя бабуся народилася 07.11.1919, у важкий післявоєнний час."
result = process_sentence(translator, tokenizer, text)
print(result)
# Output: "Моя бабуся народилася сьомого листопада тисяча дев'ятсот дев'ятнадцятого року, у важкий післявоєнний час."
```

## Performance Comparison

| Metric | Original Model | CTranslate2 Model |
|--------|---------------|-------------------|
| Inference time (avg) | ~300ms | ~70ms |
| Memory usage | ~2.5GB | ~900MB |
| Batch processing (10) | ~2.8s | ~0.7s |

*Measurements taken on NVIDIA T4 GPU with int8_float16 quantization

## Integration

This model can be easily integrated into production systems using:
- Python with CTranslate2 library
- C++ applications via CTranslate2 C++ API
- REST API services
- Docker containers

## Model Information

For detailed information about the model capabilities, training data, and usage examples, please refer to the original model: [skypro1111/m2m100-ukr-verbalization](https://huggingface.co/skypro1111/m2m100-ukr-verbalization)

## License

This model is released under the MIT License, in line with the original m2m100-ukr-verbalization model.