Enhancing Persona Consistency with Large Language Models
Haozhe Shi, Kun Niu · 2024
The capabilities of Large Language Models (LLMs) have been remarkably enhanced through constant development and display impressive capabilities in numerous Natural Language Processing (NLP) tasks. This paper introduced an effective method to utilize LLMs to extract relationships, environmental context, and storylines from character information to generate dialogue datasets that enhance the fine-tuned models’ persona consistency. Then, this research employed Low-Rank Adaptation (LoRA) to perform cost-effective fine-tuning of the LLMs based on these dialogue datasets, thus creating models that can act as the specific character. Finally, this paper evaluated the fine-tuned models using automated metrics for objective analysis and conducted an evaluation focusing on the ability of the models to maintain persona consistency, including self-awareness, robustness, and individuality, and the fine-tuned models demonstrated superior performance over baseline models.