CodingWithBashir/Bwenge-Ai

A Kinyarwanda based Ai # BwengeAi - Rwandan AI that knows Kinyarwanda BwengeAi is a Rwandan AI model trained on Kinyarwanda language data from multiple sources including Huggingface datasets, Kinyarwanda Wikipedia, and Rwandan news content. ## Features - **Multi-source data collection** from Huggingface, Wikipedia, and Igihe - **Kinyarwanda language support** with comprehensive NLP capabilities - **Efficient fine-tuning** using LoRA for parameter-efficient training - **Evaluation pipeline** with perplexity, ROUGE, BLEU, and accuracy metrics ## Project Structure ``` BwengeAi/ ├── configs/ │ └── default.yaml # Main configuration ├── src/ │ ├── data_collection/ # Data collection modules │ │ ├── huggingface_collector.py │ │ ├── wikipedia_collector.py │ │ ├── igihe_scraper.py │ │ └── data_processor.py │ ├── models/ # Model architecture │ │ └── bwenge_model.py │ ├── training/ # Training pipeline │ │ └── trainer.py │ └── evaluation/ # Evaluation metrics │ └── metrics.py ├── scripts/ │ ├── collect_all.py # Data collection script │ └── train.py # Training script ├── data/ │ ├── raw/ # Raw collected data │ └── processed/ # Processed training data ├── tests/ # Test files └── docs/ # Documentation ``` ## Installation ```bash # Clone the repository git clone cd BwengeAi # Create virtual environment python -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate # Install dependencies pip install -e ".[dev]" ``` ## Data Sources ### Huggingface Datasets - **CircuitNotion/kinyarwanda_corpus** - Large Kinyarwanda text corpus (3.52M rows) - **mbazaNLP/kinyarwanda_monolingual_v01.1** - Monolingual corpus (1.07M rows) - **saillab/alpaca_kinyarwanda_taco** - Instruction tuning (62k rows) - **mbazaNLP/Kinyarwanda_English_parallel_dataset** - Translation pairs (55.7k rows) - Plus 100+ additional datas …