与自己对话：在大型语言模型中抵御遗忘

Catastrophic forgetting poses a significant challenge when fine-tuning large language models (LLMs) on narrow, task-specific data, often degrading their general knowledge and reasoning capabilities. This study introduces SA-SFT, a lightweight self-augmentation routine where an LLM generates self-dialogues prior to fine-tuning. The resulting self-authored data is mixed with task data without altering optimization or training schedules. SA-SFT effectively mitigates catastrophic forgetting while enhancing in-domain performance, requiring no external data or additional tuning. Across 50 evaluation scenarios, it maintains performance comparable to the original model and achieves the best results in 40 cases, outperforming common baselines.