A Study on Bias Analysis of Open-Source Large Language Models(LLMs) and Prompt-Based Improvement Measures for Corporate Adoption

This paper proposes a Chain of Thought(CoT) based bias mitigation prompt technique optimized for Korean to address the bias issues that arise when companies introduce open-source based large language models(LLMs). Based on the evaluation methodology used in the existing paper and the KosBi Data Set and KcELECTRA classifiers, this study systematically analyzed the bias of the latest open-source LLM(Large Language Model) and provided the experimental results that could be considered for actual company introduction. This paper demonstrates that simple prompt example provision alone has limitations in improving bias and experimentally demonstrates that the bias of LLM can be effectively mitigated by applying CoT-style system prompts that induce a step-by-step thought process. The significance of this study is that it presented criteria for introducing reliable models and specific bias management strategies to increase the practical availability of open-source LLM. Through this, companies can come up with realistic measures to effectively manage internal and external service quality and corporate image while simultaneously securing data security and fairness.