“Google's Gemma models were trained using knowledge distillation, specifically by minimizing the KL divergence with the output distribution of the larger Gemma 27B model.”