归档 2025 09 / 12 How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Composition