저자원 환경에서의 LoRA 튜닝을 활용한 LLM 저작권 침해 응답 제어

초록

As the use of Large Language Models (LLMs) rapidly expands, concerns about copyright infringement have intensified. This study aims to fine-tune a pre-trained LLM in a low-resource environment to selectively reject responses that potentially violate copyright, all while maintaining original performance. To achieve this, we utilize Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method. We constructed a balanced Korean-based dataset, combining commonsense data with copyright refusal samples, to mitigate catastrophic forgetting and mode collapse. The proposed model achieved an accuracy of 88.67%, more than double that of the Trillion-7B base model. Furthermore, by reducing verbosity and significantly shortening response times, our resource-efficient fine-tuning approach demonstrates that effective copyright infringement prevention is achievable.

키워드