Walk me through how you would implement knowledge distillation to train a smaller student model from a large transformer.
