Explore parameter efficient fine-tuning methods on Large Language Model

Recent advancements in Large Language Models (LLMs) have enabled the development of a single model capable of performing a wide range of tasks. However, the cost of training and fine-tuning LLMs for unseen tasks is extremely high and time-consuming, rendering it inaccessible to many researchers and organizations with limited resources. Researchers have proposed various methods to encounter this issue. One proposed method is parameter-efficient fine-tuning (PEFT), which typically involves adapting only a small subset of the model parameters while preserving the model's performance. The recently proposed PEFT method IA3 has demonstrated promising results. However, there is limited research in the evaluations on IA3, prompt tuning, and model merging using PEFT vectors. This work serves as the first step toward understanding weight space in PEFT vectors and its application in model merging. Through experiments on fine-tuning T5 using IA3 and model merging using task vectors, we discovered prompts' influence on the fine-tuning results, the correlations among PEFT vectors in terms of task types, various approaches in merging models using task vectors, and generalization ability in the commonsense dataset.Bachelor of Scienc