john200tim-hash/KipsigisLLM

An language model for the Kipsigis Language. # KipsigisLLM A modular, scalable, and open-source foundation for building and training a custom Language Model (LLM) for the Kipsigis language. --- ## 📁 Repository Structure & Organization ```text KipsigisLLM/ ├── configs/ │ └── train_config.yaml # Training & model architecture configuration ├── data/ │ ├── raw/ # Raw text sources (place .txt files here) │ │ └── sample_kipsigis.txt # Provided starter dataset for instant training │ ├── processed/ # Merged & processed corpus generated by train.py │ └── datasets/ # Optional location for dataset artifacts ├── tokenizer/ │ └── kipsigis_tokenizer.json # Generated BPE tokenizer ├── checkpoints/ │ └── kipsigis_model.pt # Saved model PyTorch weights ├── src/ │ ├── data_prep.py # Data cleaning and corpus consolidation logic │ ├── dataset.py # PyTorch Dataset and DataLoader with val split │ ├── model.py # Causal Transformer (Decoder-only) LLM architecture │ └── tokenizer_train.py # BPE Tokenizer training utility ├── train.py # Main training pipeline script ├── inference.py # Text generation and evaluation script ├── run_train.sh # Unix (Linux/macOS) training runner script ├── run_inference.sh # Unix (Linux/macOS) inference runner script ├── run_train.bat # Windows training runner script ├── run_inference.bat # Windows inference runner script ├── requirements.txt # Python dependencies └── README.md ``` ### What is Tracked vs Untracked in Git? - **Tracked in Repo:** Core pipeline logic, configuration (`configs/train_config.yaml`), runner scripts, setup files, and `data/raw/sample_kipsigis.txt` for instant dry-runs. - **Excluded / Untracked (in `.gitignore`):** Proprietary text corpora (`data/raw/*` except sample data), processed corpora (`data/processed/*`), custom trained weights …