[2G1-GS-11] AI and Society:

[2G1-GS-11-02] Proposal and evaluation experiment of Fine-Tuning method using copywriter's thinking process

○Daisuke Niino¹, Yasuharu Sasaki¹, Susumu Namikawa¹, Dai Hirose¹, Yosuke Yamagami¹, Takuma Kawada¹, Kazuya Kishimoto¹, Kobori Tomoki¹, Nao Arai¹, Hiroki Kita¹, Shinya Shiragami², Masahiro Fukuda², Takuya Ito², Yoshia Abe³, Yoshimasa Tsuruoka³, Hitoshi Matsubara³ (1. DENTSU INC., 2. Dentsu Digital Inc., 3. AI Center, The Univ. of Tokyo)

Keywords:Large language Models, Fine-Tuning, Process of Thinking, Creative Task, Computer Supported Cooperative Work

When an LLM is used in creative tasks, satisfactory output is not always obtained due to a lack of expert knowledge. Therefore, we examine how fine-tuning with a copywriter’s thought process as the training data would change the quality of the output by comparing it with the output of an LLM without fine-tuning. Our experimental results show that the fine-tuned LLMs achieve a 5% higher overall rating, and the sub-item ratings of “period” and “charming” increase by 24% and 13%, respectively.