Quantification of Large Language Model Distillation