Two-stage LLM Fine-tuning with Less Specialization and More Generalization