DynamicTune

DynamicTune is a machine learning tool that enables direct closed-form weight surgery to transfer information from larger teacher transformer models to smaller student models across different architectures and dimensions without backpropagation or retraining. It improves student model accuracy while preserving language modeling integrity by aligning hidden trajectory dynamics based on spectral entropy analysis, enabling efficient model compression and transfer.