A new benchmark suite for Korean, called KBS, aims to provide a broad evaluation of foundation models in the language. The suite brings together six existing benchmark formats adapted or recreated for Korean: Ko-ARC, Ko-GSM8K, Ko-EQ-Bench, Ko-WinoGrande, Ko-LAMBADA, and Ko-IFEval. Together they cover six distinct capabilities, from reasoning and math to instruction following and language understanding.

The suite contains 9,396 items in total. According to the paper, the construction process combined review, localization, and direct Korean creation. This suggests a mix of translating existing tasks and generating new ones, though the abstract does not specify the exact split.

By offering a standardized Korean evaluation set, KBS could help researchers compare models more reliably and identify gaps in Korean language performance. The authors present it as a resource for foundation model evaluation, though details on baseline results or validation are not included in the abstract.