ManalOutmani/Translation-of-Tamazight

# Enhanced English-to-Tamazight Neural Machine Translation A character-level Seq2Seq neural machine translation model for translating English to Tamazight (Berber/Amazigh) with both Latin and Tifinagh script support. ## Overview This project implements an enhanced neural machine translation system specifically designed for the low-resource English→Tamazight language pair. The model achieves competitive performance (targeting BLEU 12-18+) through several advanced techniques including attention mechanisms, learning rate scheduling, and beam search decoding. ## Features ### Core Capabilities - **Bidirectional Translation Support**: Translates English text to Tamazight - **Dual Script Output**: Generates translations in both Latin script and Tifinagh (ⵜⵉⴼⵉⵏⴰⵖ) - **Character-Level Architecture**: Handles morphologically rich Tamazight effectively - **Low-Resource Optimizations**: Specialized techniques for limited training data ### Advanced Enhancements - **Attention Mechanism**: Improves translation quality for longer sequences - **Cosine Learning Rate Scheduling**: Includes warmup period for stable training - **Validation Split & Early Stopping**: Prevents overfitting on limited data - **Beam Search Decoding**: Explores multiple translation possibilities for better quality - **Automatic Model Checkpointing**: Saves the best performing model during training ## Installation ### Prerequisites - Python 3.8+ - PyTorch 1.9+ - CUDA (optional, for GPU acceleration) ### Required Packages ```bash pip install torch torchvision torchaudio pip install datasets pip install numpy ``` ### Optional (for evaluation) ```bash pip install sacrebleu # For BLEU score calculation pip install huggingface-hub # For dataset authentication ``` ## Dataset The model trains on two Hugging Face datasets: 1. **Weblate-Translations** (`Tamazight-NLP/Weblate-Translations`) - No authentication required - Software localization translations 2. **Beni-Mellal-Tamazight** (`Tamazight-NLP/Beni- …