Science ministry opens 29 AI data sets from elite model project

The science ministry on Thursday began opening 29 types of artificial intelligence training data secured by five elite teams under its proprietary AI foundation model project. The release includes approximately 35.44 million items amounting to an estimated 1.56 trillion tokens.

Consortia led by Naver Cloud, Upstage, SK Telecom, NCAI and LG AI Research built the data sets using a budget allocated in 2025. Contents range from massive pre-training collections and multimodal video or voice records to red-teaming materials designed to ensure model safety.

Naver Cloud contributed 2.34 million public videos and 520,000 broadcast clips alongside text captions and voice question-and-answer pairs, while Upstage provided data for foundational learning and items for agent development. SK Telecom focused on high-difficulty reasoning tasks in math, science and law, whereas NCAI supplied manufacturing documents and customer service logs to enhance long-context understanding. LG AI Research added a physical AI data set filmed in 50 actual Korean homes to support vision and robotics applications.

The National Information Society Agency and the Telecommunications Technology Association verified the quality and screened out personal information before making the files available for free download on the AI Hub website.

Kim Kyung-man, head of the ministry's AI policy office, said the open data represents valuable assets containing the strategic know-how of the elite teams. He added that the initiative aims to foster qualitative growth across the entire AI ecosystem rather than merely supporting individual model development.