Kakao Reveals Kanana LLM Post-training Process… Overwhelming Korean Language Performance

Kakao's Kanana LLM applied Chat Templates to respond according to user intent and performed Supervised Fine-tuning (SFT) using data from various domains, including General-chat, Code, and Math. Subsequently, both Offline (DPO) and Online Preference-based Learning were applied to strengthen the model's preference alignment, achieving optimal performance through Model Merging techniques at each stage. Notably, in benchmark evaluations such as LogicKor and MT-Bench, the Kanana Essence model showed performance comparable to similar-sized open-source models in English and general conversation capabilities, while recording significantly higher scores than other models in Korean evaluations. The Kanana Nano model also demonstrated an even larger advantage over the Essence model in Korean performance, highlighting the necessity of developing proprietary models based on high-quality Korean data.