Faster Translate

A high-performance translation library powered by state-of-the-art models. Faster Translate offers optimized inference using CTranslate2 and vLLM backends, providing an easy-to-use interface for applications requiring efficient and accurate translations.

🚀 Features

• High-performance inference using CTranslate2 and vLLM backends
• Seamless integration with Hugging Face models
• Flexible API for single sentence, batch, and large-scale translation
• Dataset translation with direct Hugging Face integration
• Multi-backend support for both traditional (CTranslate2) and LLM-based (vLLM) models
• Text normalization for improved translation quality

📦 Installation

pip install faster-translate

Optional Dependencies

For specific normalizers or model backends:

# For Bengali text normalization
pip install git+https://github.com/csebuetnlp/normalizer

# For vLLM backend support (required for LLM-based models)
pip install vllm

🔍 Usage

Basic Translation

from faster_translate import TranslatorModel

# Initialize with a pre-configured model
translator = TranslatorModel.from_pretrained("banglanmt_bn2en")

# Translate a single sentence
english_text = translator.translate_single("দেশে বিদেশি ঋণ নিয়ে এখন বেশ আলোচনা হচ্ছে।")
print(english_text)

# Translate a batch of sentences
bengali_sentences = [
    "দেশে বিদেশি ঋণ নিয়ে এখন বেশ আলোচনা হচ্ছে।",
    "রাত তিনটার দিকে কাঁচামাল নিয়ে গুলিস্তান থেকে পুরান ঢাকার শ্যামবাজারের আড়তে যাচ্ছিলেন লিটন ব্যাপারী।"
]
translations = translator.translate_batch(bengali_sentences)

Using Different Model Backends

# Using a CTTranslate2-based model
ct2_translator = TranslatorModel.from_pretrained("banglanmt_bn2en")

# Using a vLLM-based model
vllm_translator = TranslatorModel.from_pretrained("bangla_qwen_en2bn")

Loading Models from Hugging Face

# Load a specific model from Hugging Face
translator = TranslatorModel.from_pretrained(
    "sawradip/faster-translate-banglanmt-bn2en-t5",
    normalizer_func="buetnlpnormalizer"
)

Translating Hugging Face Datasets

Translate an entire dataset with a single function call:

translator = TranslatorModel.from_pretrained("banglanmt_en2bn")

# Translate the entire dataset
translator.translate_hf_dataset(
    "sawradip/bn-translation-mega-raw-noisy", 
    batch_size=16
)

# Translate specific subsets
translator.translate_hf_dataset(
    "sawradip/bn-translation-mega-raw-noisy",
    subset_name=["google"], 
    batch_size=16
)

# Translate a portion of the dataset
translator.translate_hf_dataset(
    "sawradip/bn-translation-mega-raw-noisy",
    subset_name="alt",
    batch_size=16, 
    translation_size=0.5  # Translate 50% of the dataset
)

Publishing Translated Datasets

Push translated datasets directly to Hugging Face:

translator.translate_hf_dataset(
    "sawradip/bn-translation-mega-raw-noisy",
    subset_name="alt",
    batch_size=16, 
    push_to_hub=True,
    token="your_huggingface_token",
    save_repo_name="your-username/translated-dataset"
)

🌐 Supported Models

🛠️ Advanced Configuration

Custom Sampling Parameters for vLLM Models

from vllm import SamplingParams

# Create custom sampling parameters
sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=512
)

# Initialize translator with custom parameters
translator = TranslatorModel.from_pretrained(
    "bangla_qwen_en2bn", 
    sampling_params=sampling_params
)

💪 Contributors

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📚 Citation

If you use Faster Translate in your research, please cite:

@software{faster_translate,
  author = {Sawradip Saha and Contributors},
  title = {Faster Translate: High-Performance Machine Translation Library},
  url = {https://github.com/sawradip/faster-translate},
  year = {2024},
}