Cost-effective Distillation of Large Language Models