KETI released four Korean multimodal AI datasets. K-MIND contains 115.6 hours of speech from 286 speakers; KoMMERCE includes 15,293 shopping conversations. KPrism offers 1,002 multimodal hallucination-evaluation questions, while KoEVD contains 13,552 utterances and 71,395 responses. Distribution includes AI Hub and Hugging Face, but the datasets have different training and evaluation purposes. Access alone does not guarantee Korean-language accuracy or eliminate model hallucinations.
Back to Tech
KETI releases four Korean multimodal AI datasets
Training conversations and hallucination evaluations serve different purposes and conditions.