MNIST digit classification using DSPy framework with support for various LLM backends. Configure optimizers and training iterations for optimal performance.
output.mp4
- DSPy-powered MNIST classification with support for multiple LLM backends
- Flexible optimizer selection (MIPROv2, BootstrapFewShot)
- Configurable training iterations and worker threads
- Model selection (supports any LLM compatible with DSPy)
- Response caching for faster iterations
git clone https://github.com/tom-doerr/mnist_dspy.git
cd mnist_dspy
pip install -r requirements.txtmnist_dspy/
├── mnist_trainer.py # Main training script with optimizer selection
├── mnist_data.py # MNIST data loading and preprocessing
└── mnist_dspy.py # DSPy model definitions
Train a model with specific optimizer and iterations:
# Use MIPROv2 optimizer with light optimization
python mnist_trainer.py --optimizer MIPROv2 --iterations 1 --model your-llm-model --auto light
# Use MIPROv2 with medium optimization and caching disabled
python mnist_trainer.py --optimizer MIPROv2 --iterations 1 --model your-llm-model --auto medium --no-cache
# Use BootstrapFewShot optimizer with custom number of workers
python mnist_trainer.py --optimizer BootstrapFewShot --iterations 1 --model your-llm-model --num-workers 50--optimizer: Choose between 'MIPROv2' or 'BootstrapFewShot' (default: MIPROv2)--iterations: Number of optimization iterations (default: 1)--model: Model identifier for your LLM--auto: Optimization intensity for MIPROv2 ['light', 'medium', 'heavy'] (default: light)--num-workers: Number of worker threads (default: 100)--no-cache: Disable LLM response caching
The trainer:
- Uses a subset of MNIST data for faster experimentation
- Processes 10,000 training examples
- Evaluates on 200 test examples
- Supports multi-threaded evaluation
Distributed under the MIT License. See LICENSE for more information.