Kakao's 'Kanana-2' surpasses global models in Korean AI safety evaluation

Kakao has released the safety verification results for its self-developed lightweight small language model (SLM) series, 'Kanana-2'. The evaluation focused on the safety of the models in handling Korean language AI tasks.

The models made available on the open-source platform Hugging Face on August 28, namely 'Kanana-2-1.3B-Instruct' and 'Kanana-2-3B-Instruct', were the subjects of this assessment. These models are designed to be efficient and accessible while maintaining high performance standards.

In a comparative analysis, both 'Kanana-2' models achieved top rankings. They were compared against models with similar parameter sizes, including Google's 'Gemma' and Alibaba's 'Qwen'. The results showed that Kakao's models secured the overall first place in the evaluations, surpassing their global counterparts.

This achievement highlights Kakao's advancements in developing safe and reliable AI language models, particularly for the Korean language. The success of 'Kanana-2' suggests a competitive edge for Kakao in the rapidly evolving field of artificial intelligence.