A lightweight CLI for simple model preprocessing, training, evaluation, and optimization.

nano-llm is a streamlined command-line interface for simple transformer model development. It provides a complete workflow from data preprocessing to model training, evaluation, and text generation, designed for educational purposes and small-scale experiments.

• Tokenizer Training: Train custom tokenizers on your text data
• Model Training: Train transformer-based language models
• Model Evaluation: Evaluate models with perplexity metrics
• Text Generation: Generate text using trained models
• YAML Configuration: Simple configuration-based approach

• Model Optimization: Implement pruning and distillation techniques
• Advanced Positional Encoding: Add Rotary positional encoding (RoPE) for better performance
• Efficient Normalization: Replace LayerNorm with RMSNorm for improved efficiency
• HuggingFace Integration: Use the transformers package to enable model deployment to HuggingFace Hub

# Clone the repository
git clone https://github.com/ssubedir/nano-llm.git
cd nano-llm

# Install dependencies
uv sync

For a complete workflow example, see the Getting Started Guide.

For detailed information about commands, configuration, and examples, see the docs folder:

• Getting Started - Complete workflow tutorial
• Configuration Reference - All configuration options
• Command Guides - Detailed command documentation:
  • Tokenizer
  • Training
  • Evaluation
  • Generation

nano-llm/
├── app/                   # Core application code
│   ├── cli/               # Command-line interface
│   ├── data/              # Data processing
│   └── model/             # Model architecture
├── configs/               # Configuration files
├── dataset/               # Sample datasets
└── docs/                  # Documentation

• Python 3.12+
• PyTorch
• CUDA-compatible GPU (recommended for training)

# Lint all files in the current directory
uvx ruff check

# Format all files in the current directory
uvx ruff format

Contributions are welcome!

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Run linting and formatting:

   uvx ruff check
   uvx ruff format

5. Submit a pull request

See the TODO section for areas that need work.

MIT License - see the LICENSE file for details.