ndagirenairah/ENGLISH-LUGANDA-TRANSLATOR

# English-Luganda Translator A machine learning translator for English ↔ Luganda using transformer sequence-to-sequence models. **Status**: [SUCCESS] Clean ML pipeline ready for training on Google Colab GPU --- ## [SUMMARY] Overview This project implements a complete machine learning workflow for English-Luganda translation: - **Model**: Helsinki-NLP/OPUS-MT-en-mul (143.1M parameters) - **Framework**: PyTorch + HuggingFace Transformers - **Training**: Google Colab with Tesla T4 GPU (or CPU fallback) - **Performance**: BLEU score 20.69 on 5,008 test samples - **Visualizations**: Data distribution, training results, quality metrics dashboard --- ## [STATS] Datasets The project uses **5 real English-Luganda parallel datasets**: 1. **Kambale Corpus** (~50,012 pairs) - High-quality parallel translations - Domain: Agriculture, society, community 2. **Cultural Dataset** (~12 pairs) - Buganda cultural and heritage terms - Domain: Traditions, kingship, culture 3. **JW300** (~15 pairs) - Religious and literary texts - Domain: Spiritual, philosophical 4. **Makerere NLP** (~15 pairs) - Academic Luganda - Domain: Formal, educational 5. **Sunbird SALT** (~18 pairs) - Low-resource language data - Domain: General language patterns **Total**: 50,072 real translation pairs **Data Splits**: - Training: 40,056 pairs (80%) - Validation: 5,008 pairs (10%) - Testing: 5,008 pairs (10%) --- ## 🏗️ Project Structure ``` src/ # ML Pipeline modules ├── config.py # Centralized configuration ├── utils.py # Helper functions ├── 1_load_data.py # Step 1: Load all 5 datasets ├── 2_preprocess.py # Step 2: Create train/val/test splits ├── 3_train.py # Step 3: Train transformer model └── 4_evaluate.py # Step 4: Evaluate & calculate BLEU scripts/ └── run_pipeline.py # Run complete pipeline data/ ├── raw/ # Original CSV files (5 datasets) │ ├── kambale_train.csv │ ├── cul …