Nemotron-Personas-Korea: AI Takes Root in Korean Culture

The artificial intelligence (AI) models that power the majority of agents today are primarily trained on English web data. This poses a problem when it comes to adapting to the cultural and linguistic specifics of other countries, such as South Korea. Indeed, these models do not take into account Korean honorific structures, regional occupation patterns, and the cultural context that Korean users expect. For example, an agent applying American healthcare workflows to the Korean public health system is not ready for production and may not adequately meet the needs of Korean users.

To address this situation, Nemotron-Personas-Korea has been created. This dataset provides 6 million fully synthetic personas, grounded in official statistics and reference data from sources such as the Korean Statistical Information Service (KOSIS), the Supreme Court of Korea, the National Health Insurance Service, and the Korea Rural Economic Institute. NAVER Cloud also contributed to this project by providing reference data and domain expertise during the design phase.

Each persona is demographically accurate but contains no personally identifiable information (PII), thus complying with the Korean Personal Information Protection Act (PIPA). South Korea is also one of the few countries to publish an official guide on generating synthetic data, establishing rules for anchoring models with synthetic versions of sensitive data. This dataset follows this rigorous approach.

In this tutorial, we will transform a synthetic persona into a deployed Korean agent—from filtering the dataset to inference—in about 20 minutes using hosted APIs.

A Sovereign Dataset for South Korea

The Nemotron-Personas-Korea dataset is vast and detailed. It consists of 7 million records, each containing 26 different fields. These fields include 7 persona fields, 6 persona attribute fields, 12 demographic and geographic contextual fields, and 1 unique identifier. The geographic coverage is comprehensive, encompassing all 17 Korean provinces and 25 districts. In terms of name diversity, the dataset contains approximately 209,000 unique names, including 118 surnames and about 21,400 first names.

In addition to geographic and nominal diversity, the dataset covers over 2,000 occupational categories, reflecting various sectors such as technology, manufacturing, the public sector, as well as fields like sports, arts, travel, and cuisine. The personas include varied profiles such as students, military personnel, employees, unemployed individuals, and retirees.

Nemotron-Personas-Korea was generated using the NeMo Data Designer, an open-source composite AI system from NVIDIA for synthetic data. The pipeline combines a Probabilistic Graphical Model (Apache-2.0) for statistical anchoring with Gemma-4-31B for generating narratives in the Korean language. Population data comes from KOSIS, covering versions from 2020 to 2026, while name distributions come from the Supreme Court of Korea.

This dataset is the latest addition to the Nemotron-Personas Collection, which also includes datasets for the United States, Japan, India, Singapore (in collaboration with AI Singapore), Brazil (with WideLabs), and France (with Pleias). For those building a multilingual agent serving Korean users alongside other markets, it is possible to mix personas from different countries within the same pipeline.

Why This Matters for Autonomous Agents

Most agents today are blind to identity. They follow instructions without any anchor on whom they are serving. For example, an agent making an appointment at a Korean hospital using American scheduling conventions, or addressing a 60-year-old patient using 반말 (“banmal,” informal language), does not seem right. This fails.

Nemotron-Personas-Korea changes this by giving your agent a Korean operational context. Load a persona into the system prompt, and the agent inherits the region, occupation, communication norms, and domain expertise of that persona.

This works in any agent framework. Deploy with NemoClaw (NVIDIA's open-source reference stack for always-on agents operating in NVIDIA OpenShell environments, on everything from RTX PCs to DGX Spark), serve via NVIDIA NIM for production inference, or call the NVIDIA API directly. The persona layer is framework-independent, acting as a well-structured system prompt anchored in real Korean demographic data.

Tutorial: From Synthetic Persona to Sovereign Agent

The process of creating a Korean AI agent with Nemotron-Personas-Korea begins with loading and exploring the dataset. Developers can filter personas by occupation, region, or age to find those that match their target domain. Then, they define the agent's behavior using the structured data from the personas, allowing the agent to reason like a Korean professional.

Finally, the agent can be deployed using various options, such as the NVIDIA API Catalog, NVIDIA NIM for self-hosted inference, or NemoClaw for deploying always-on agents.

What Anchoring Changes

Here’s the same question — "독감 예방접종은 언제 맞아야 하나요?" (When should I get vaccinated against the flu?) — answered with and without persona anchoring.

Without Personas

• Responds in generic English/Korean
• References CDC/global guidelines
• "Visit your local clinic"

With Personas of Korean Healthcare Workers

• Responds in appropriate 존댓말 natural for a health consultation
• References the schedule of Korean 보건소, national vaccination program
• "가까운 보건소에서 무료 접종이 가능합니다" with regional context
• Cites Korean public health policy, uses professional medical Korean

The persona goes beyond translation — it contextualizes and results in an agent that your users will trust.

Come Build with Us in Seoul

The NVIDIA Nemotron Developer Days are taking place in Seoul today and tomorrow, April 21-22, 2026 — the first time this event is held outside of GTC. Two days of activities, including technical sessions on sovereign AI and open models, as well as a hands-on hackathon where you will have the opportunity to use Nemotron-Personas-Korea to build domain-specific Korean agents and a claw. 🦞

Join in person or via livestream. Share what you build for a chance to be featured in a future NVIDIA tutorial.