NVIDIA Releases 7 Million Synthetic Korean Personas Dataset for AI Agent Localization

NVIDIA released Nemotron-Personas-Korea, a dataset containing 7 million demographically accurate synthetic personas grounded in official Korean government statistics. The dataset addresses a core problem: AI agents trained primarily on English data lack the cultural context, honorific structures, and regional patterns needed for Korean production deployments.

Dataset Specifications

The dataset provides 1 million base records, each containing 7 distinct personas, for a total of 7 million personas. Each persona includes 26 structured fields: 7 persona fields, 6 persona attribute fields, 12 demographic and geographic contextual fields, and 1 unique identifier.

Geographic coverage spans all 17 Korean provinces and 25 districts. Names draw from approximately 209,000 unique combinations (118 surnames, roughly 21,400 given names). Occupations cover 2,000+ categories reflecting Korea's tech, manufacturing, and public sectors.

Persona types include professional, family, sports, arts, travel, culinary, and concise variants. Life stages cover students, military service, employed, unemployed, and retired individuals. All narrative content is in natural Korean.

Data Sources and Generation

Source data comes from the Korean Statistical Information Service (KOSIS) 2020-2026 releases, Supreme Court of Korea name distributions via namechart.kr, National Health Insurance Service records, and the Korea Rural Economic Institute. NAVER Cloud contributed seed data and domain expertise during design.

NVIDIA generated the dataset using NeMo Data Designer, pairing a Probabilistic Graphical Model (Apache-2.0) for statistical grounding with Gemma-4-31B for Korean-language narrative generation. The dataset is released under CC BY 4.0 license.

The dataset contains zero personally identifiable information and was designed with Korea's Personal Information Protection Act (PIPA) in mind. South Korea is among the few countries with an official Synthetic Data Generation guide, and this dataset follows that governance approach.

Integration With Agent Frameworks

The dataset integrates with NVIDIA's agent deployment stack. Developers can deploy using NemoClaw (NVIDIA's open-source reference stack for always-on agents), serve through NVIDIA NIM for production inference, or call the NVIDIA API directly. The persona layer acts as structured system prompts and is framework-agnostic.

NVIDIA demonstrated a 20-minute workflow: filter the dataset by occupation and region, extract structured fields (name, region, occupation, skills), construct a Korean-language system prompt with behavioral guidelines, and connect to inference via NVIDIA API catalog or self-hosted NIM deployments.

Example use cases include healthcare agents that understand Korean public health workflows, financial advisors grounded in Korean banking systems, and education assistants that use appropriate honorific structures based on age and social context.

Nemotron-Personas Collection

Nemotron-Personas-Korea joins NVIDIA's broader Nemotron-Personas Collection, which 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 in the same pipeline.

What This Means

This dataset addresses a genuine gap in localized AI deployment. Most foundation models lack grounding in non-English demographic patterns, regulatory frameworks, and communication norms. Synthetic persona datasets offer a PII-free method to inject that context into agent system prompts. The CC BY 4.0 license and reliance on official government statistics give the dataset credibility for production use cases where demographic accuracy matters — healthcare, finance, government services. The approach is replicable: pair a probabilistic model with local statistics and a strong language model, and you can generate similar datasets for other markets. NVIDIA's multi-country collection suggests this becomes infrastructure for sovereign AI systems that need to operate within specific cultural and regulatory contexts.