Understanding the Instruction Mixture for Large Language Model Fine-tuning