Adding Character to The Utterances of AIWolf Agent by Fine-Tuning GPT Language Model

Abstract

In the natural language division of 2023 International AIWolf Competition, many AI agents using LLM-based generative AI played werewolf games with more natural conversation than before. However, there are many inconsistencies between statements and between statements and actions. One possible improvement method is an architecture in which a protocol agent capable of consistent behavior and utterance wrapped in a natural language I/O part which mutually converts between AIWolf protocol and natural language. In this study, we attempted to add character to the Japanese utterances corresponding to the AIWolf protocols by applying a sentence generation method using japanese-gpt-neox-3.6b Japanese GPT language model. In order to add character to the utterances, we fine-tuned original model using the sentences extracted from “Werewolf BBS” game log data. It is confirmed that, to some extent, the obtained model can generate sentences that are characterized as spoken by an old man.