KULLM3, LLM Tailored for Korean Language, Shows Remarkable Advances in Language Generation
NLP&KULLM3, the Korean-specialized LLM developed through collaboration between NLP & AI Lab and HIAI Research Center, has been unveiled.
Led by Professor Lim Heui-seok from the Department of Computer Science, the NLP & AI Lab and HIAI Research Center first introduced KULLM in June 2023, with the aim of enhancing Korean language generation capabilities to a level practical for real-world use.
To achieve this, considerable effort was put into creating high-quality Korean instruction datasets, which significantly improved the ability to follow Korean instructions. The research team produced various Korean datasets related to different desks and even created specialized data specifically for the KULLM3 model. This high-quality data was applied to Upstage’s SOLAR-10.7B model through instruction tuning learning, resulting in the birth of KULLM3.
Analysis conducted by the research team using GPT-4 Turbo revealed that the generated responses of KULLM3 were superior to those of existing Korean models, exhibiting capabilities comparable to GPT-3.5 Turbo and GPT-4 Turbo. According to the performance metrics provided by the research team, the model demonstrated excellent performance in evaluations of response fluency, coherence, accuracy, completeness, and overall quality.
The research team emphasized the excellent ability of the KULLM3 model to understand and execute instructions in Korean, expecting it to be utilized in various fields such as AI counseling chatbots and RAG-based question-answering systems.
Professor Lim from the Department of Computer Science said, “There is currently a lot of interest in building private LLMs and on-premise LLMs for the Korean language.” He added, “We expect KULLM3, with its significantly improved Korean language generation performance, to serve as a good alternative.”
NLP&NLP & AI Lab and HIAI Research Center annually present top-tier papers at the most prestigious conferences in the field of natural language processing and are at the forefront of research on everyday AI utilizing LLMs, making them leaders in the domestic artificial intelligence sector.