Behind the Secrets of Large Language Models