An Empirical Study of Instruction-tuning Large Language Models in Chinese