[3Xin2] Poster session 1

[3Xin2-08] Effect of Fine-Tune and Few Shot Prompt in Character Profile for LLM Dialog Generation

〇Nanami Ozawa¹, Kano Yoshinobu¹ (1.Shizuoka University)

Keywords:Generation AI, Fine Tuning, Character

The advent of Large Language Models (LLMs) has brought significant attention to conversational AI. In this research, we investigated how training LLMs with different amounts of data and quality, using methods such as Prompt Engineering, Supervised Fine Tuning (SFT), and Reinforcement Learning from Human Feedback (RLHF), can generate dialogues that do not deviate from character settings. We also experimented with generating responses when given content that contradicts the character settings used in the learning data for SFT and RLHF. The results confirmed that it is possible to produce the expected names from within and outside the learning data. It was observed that using RLHF allows for the output of names in the learning data without deviating from the character settings. When provided with content that contradicts the learning data used in SFT or RLHF, the influence of character settings and Few Shot Prompts was suggested to be more significant than the learning from SFT or RLHF.