NVIDIA releases Nemotron-Personas-Korea, a 7-million-record synthetic dataset for grounding AI agents in Korean demographics and language norms

NVIDIA has released Nemotron-Personas-Korea, a 7-million-record synthetic dataset designed to ground AI agents in authentic Korean demographics and cultural context. The dataset covers all 17 Korean provinces and includes 2,000+ occupation categories, ~209,000 unique names reflecting actual Korean naming distributions, and 26 structured fields spanning persona types, life stages, skills, and geographic context.

The personas are built from official seed data: population statistics from the Korean Statistical Information Service (KOSIS) covering 2020–2026, name distributions from the Supreme Court of Korea, healthcare context from the National Health Insurance Service, and rural economic expertise from the Korea Rural Economic Institute. NAVER Cloud contributed additional domain knowledge during dataset design. Critically, every persona contains zero personally identifiable information (PII), meeting the requirements of South Korea's Personal Information Protection Act (PIPA).

The dataset was generated using NVIDIA's open-source NeMo Data Designer system, which pairs a Probabilistic Graphical Model for statistical grounding with Gemma-4-31B for natural Korean-language narrative synthesis. This approach ensures demographic fidelity while maintaining privacy compliance. NVIDIA released the dataset under a CC BY 4.0 license, making it available for training and fine-tuning workflows.

Agents built with these personas inherit region-specific knowledge, occupation context, communication norms including Korean honorifics (formal 존댓말 speech), and domain expertise. In practice, this means a public health agent can guide users through Korean clinic scheduling and insurance procedures rather than applying U.S. healthcare workflows—a distinction that separates working deployments from non-functional systems. The persona layer acts as a structured system prompt and is framework-agnostic, compatible with any agent platform.

Nemotron-Personas-Korea is the latest addition to NVIDIA's broader Nemotron-Personas Collection, which already includes datasets for the USA, Japan, India, Singapore (with AI Singapore), Brazil (with WideLabs), and France (with Pleias). Developers building multilingual agents can blend personas across countries within the same pipeline. Deployment options include NVIDIA's API catalog, self-hosted NVIDIA NIM inference servers, or NemoClaw, NVIDIA's open-source reference stack for always-on agents running on hardware ranging from RTX PCs to DGX Spark clusters.