A 1.5 million parameter stylable-level language model trained from scratch to generate text in the style of Nam Cao. The entire training stack was implemented in Go without relying on machine learning frameworks.
- Custom RNN and LSTM implementations
- Adam optimizer with automatic differentiation
- Hand-written
float32math kernels for improved CPU performance - SIMD-accelerated matrix operations
- Model checkpoint save/load support for training continuation and inference
- End-to-end training pipeline built from scratch