Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning

• Published in: arXiv.org
• Main Authors: Yang, Zhaorui, Liu, Qian, Pang, Tianyu, Wang, Han, Feng, Haozhe, Zhu, Minfeng, Chen, Wei
• Format: Paper Journal Article
• Language: English
• Published: Ithaca Cornell University Library, arXiv.org 21.02.2024
• Subjects: Computer Science - Computation and Language; Datasets; Distillation; Large language models; Natural language processing
• ISSN: 2331-8422
• DOI: 10.48550/arxiv.2402.13669

The surge in Large Language Models (LLMs) has revolutionized natural language processing, but fine-tuning them for specific tasks often encounters challenges in balancing performance and preserving general instruction-following abilities. In this paper, we posit that the distribution gap between task datasets and the LLMs serves as the primary underlying cause. To address the problem, we introduce Self-Distillation Fine-Tuning (SDFT), a novel approach that bridges the distribution gap by guiding fine-tuning with a distilled dataset generated by the model itself to match its original distribution. Experimental results on the Llama-2-chat model across various benchmarks demonstrate that SDFT effectively mitigates catastrophic forgetting while achieving comparable or superior performance on downstream tasks compared to the vanilla fine-tuning. Moreover, SDFT demonstrates the potential to maintain the helpfulness and safety alignment of LLMs. Our code is available at \url{https://github.com/sail-sg/sdft}.