A character-level language diffusion model for text generation. The model is a modified version of the nanochat gpt implementation and is trained on Tiny Shakespeare! It is only 10.7 million parameters, so you can try it out locally!

# Clone the repository
git clone <repository-url>
cd tiny-diffusion

# Install dependencies (Python 3.10+)
uv sync

The file training.py puts the weights in weights/diffusion_model.pt. The sample and animation files load the model from this file.

Currently, the weights are already provided for you! It took me around half an hour to train this model for 20,000 steps on 4xA100s. But if you want to retrain the model again, run:

# Train from scratch on Shakespeare
uv run training.py

# Training will save checkpoints to weights/diffusion_model.pt

To generate a continuous stream of output (currently 30 context lengths), run:

# Generate samples using the pre-trained model
uv run sample.py

To see the diffusion process as a nice animation, run:

# Watch the denoising process step-by-step
uv run animations/diffusion-process.py

# See Game of Life-inspired sampling (fun little experiment)
uv run animations/game-of-life.py

• Parameters: 10.7 million
• Layers: 6
• Attention Heads: 6
• Embedding Dim: 384
• Sequence Length: 256 characters
• Diffusion Steps: 128

tiny-diffusion/
├── model.py                    # Core diffusion transformer
├── training.py                 # Training script
├── sample.py                   # Text generation
├── data/
│   └── tiny_shakespeare.txt    # Training data
├── weights/
│   └── diffusion_model.pt      # Pre-trained weights
└── animations/
    ├── diffusion-process.py    # Denoising visualization
    └── game-of-life.py         # Game of Life sampling