Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning
This addresses the problem of balancing performance and preserving general abilities in fine-tuning large language models for specific tasks, representing an incremental improvement over existing methods.
The paper tackles the distribution gap between task datasets and large language models during fine-tuning, which causes issues like catastrophic forgetting, by introducing Self-Distillation Fine-Tuning (SDFT). The result shows that SDFT effectively mitigates catastrophic forgetting while achieving comparable or superior performance on downstream tasks compared to vanilla fine-tuning, as demonstrated on the Llama-2-chat model across various benchmarks.
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 https://github.com/sail-sg/sdft.