Korea University Unveils ‘Gureum3’, an LLM with Excellent Korean Language Response Ability

Excellent Ability to Understand and Execute Korean Instructions
Expectations for Various Applications Including AI Consultation Chatbots

A research team at Korea University has introduced a large language model (LLM) specialized for the Korean language called ‘Gureum3 (KULLM3)’.

Professor Lim Hee-seok, Department of Computer Science, Korea University.
Photo by Korea University

‘Gureum3’ is a Korean language-specialized LLM developed through collaboration between the NLP&AI Lab, led by Professor Heeseok Lim of the Department of Computer Science at Korea University, and the HIAI Research Institute.

Since first unveiling ‘Gureum (KULLM)’ in June last year, the NLP&AI Lab and HIAI Research Institute have been conducting research aimed at enhancing Korean language generation capabilities to a practical level.

The research team significantly improved the model’s ability to follow Korean language instructions by utilizing a high-quality Korean instruction dataset. They also created various Korean datasets and special data exclusively for the ‘Gureum3’ model.

The team explained that by applying this data to the ‘Upstage SOLAR-10.7B model’ and performing instruction-tuning training, ‘Gureum3’ was born.

Korean language-specialized LLM 'Gureum3' provided answers in Korean.
[Photo by Korea University]

The research team evaluated that “Based on analysis using GPT-4 Turbo, the generated responses of ‘Gureum3’ were far superior to existing Korean models and demonstrated capabilities comparable to GPT-3.5 Turbo and GPT-4 Turbo.” According to the performance presented by the team, the answers generated by ‘Gureum3’ showed excellent performance in fluency, coherence, accuracy, completeness, and overall score evaluations.

Professor Lim stated, “We expect that ‘Gureum3,’ which has outstanding practical Korean language generation performance, can be a good alternative.”