The Korea Advanced Institute of Science and Technology (KAIST) and Nvidia plan to make at least one million Korean-language AI training samples public in 2027 to address a data shortage.
The Korea Herald reported the plan after lab head Kim Hyun-woo, an assistant professor at KAIST, spoke with reporters in Seoul on Tuesday. Kim did not give a specific publication date.
The Korea Herald reported that Nvidia and KAIST announced plans for the lab in July. Separately, the outlet described the backing arrangement as a five-year collaboration valued at $300 million, including $250 million in computing resources supplied through Nvidia’s Korean cloud partners. The program also plans to support at least 10 KAIST researchers each year, with internship opportunities at Nvidia.
Kim said South Korea has invested heavily in AI models and computing infrastructure but devoted less attention to high-quality training material. He said evaluation resources are more available than Korean-language data for training, which still trails material in English and Chinese. He pointed to translated text and awkward phrasing in AI-generated Korean as quality concerns. Kim also warned that synthetic data without careful filtering could harm model performance.
Researchers plan to create plausible documents from limited inputs, then check them for accuracy, variety and repeated content. Kim said selecting useful samples and checking their quality take more effort than generating them. He cautioned that the initial one million samples would not by themselves close the Korean-language data gap.
The resulting datasets and methods are intended for wider use by South Korean AI developers. Kim said building South Korea's research capacity took priority over barring access by overseas competitors.
An earlier Nvidia initiative, Nemotron-Personas-Korea, used South Korean government statistics to generate seven million synthetic personas. Kim said LG AI Research used the dataset while developing K-Exaone 2.5 and that it had received about 140,000 downloads.
After building the Korean-language data foundation, the lab plans to move into industry-specific datasets and agentic AI. The Korea Herald reported that Nvidia has begun recruiting full-time researchers in South Korea, with initial hires expected by year-end. The lab has yet to settle its industry partnerships or intellectual property terms.
