Instruction Fine-Tuning Guidance: How PEFT Methods Impact the Generation of Language Model via Different Attributions

In this paper, we investigate the impact of different Parameter Efficient Fine-tuning (PEFT) methods on the generation performance of large language models, focusing on how different methods affect the model through various attributes. We evaluate five well-known PEFT methods: LoRA, AdaLoRA, LoHA, IA3, and P-Tuning taking computational resource, scalability, efficiency, and other considerations into consideration. We use the Qwen-7B-Chat model as a baseline, instruction fine-tune it using the alpaca-cleaned dataset, and use the MT-Bench and HelpSteer datasets to evaluate the model by several metrics, including helpfulness, correctness, consistency, complexity and verbosity. Experimental results show that while all methods enhance generation performance of the model, they exhibit different advantages and trade-offs in terms of text quality, diversity, and computational efficiency. Ablation studies further explore the impact of the number of training parameters and target modules on model generation performance, and find that reducing the number of them can still lead to improved performance. Our analysis provides valuable guidance for selecting the most appropriate fine-tuning methods based on specific task requirements, and enlightening thoughts on how these methods affect language model generation after instruction fine-tuning.