Improved Supervised Fine-Tuning for Large Language Models to Mitigate Catastrophic Forgetting