FPT has announced a massive expansion of Nemotron-Personas-Vietnam, its open dataset of synthetic personas designed to reflect Vietnam's official demographic and labor statistics. Developed in collaboration with NVIDIA, the dataset has grown to contain 1,000,000 synthetic personas. This tenfold increase provides developers, researchers, and enterprises with a larger and more representative foundation for building AI models, agents, and applications that understand Vietnamese people, language, and culture.

First released in June 2026, the dataset is built on NVIDIA's Nemotron-Personas methodology. This approach anchors persona generation in real-world statistics and uses open-weight large language models (LLMs) to write high-fidelity persona narratives. The resulting dataset gives developers realistic Vietnamese user profiles spanning age, region, occupation, income, education, and cultural context, all without containing any personally identifiable information (PII).

Grounding AI in Vietnam's Demographics

The Nemotron-Personas-Vietnam dataset is grounded in official statistics from Vietnam's General Statistics Office (GSO). This includes data from the Population and Housing Census 2024 and the Vietnam Household Living Standards Survey (VHLSS) 2024. Additional data contributions come from FPT Smart Cloud and the Quantum AI & Cyber Security Institute under FPT Corporation.

The dataset's structure includes:

  • 21 fields, consisting of 6 persona fields and 15 contextual fields covering persona attributes and demographics.
  • 13,000 unique full-name combinations.
  • 20 occupational categories derived from labor survey data.

By utilizing these parameters, the dataset allows developers to simulate realistic user interactions and behaviors that align with the actual demographic makeup of the country.

Technical Advantages of the Tenfold Expansion

The expansion to 1,000,000 personas represents a significant qualitative shift in how developers can train and evaluate AI models for the Vietnamese market. The scale of the updated dataset introduces several key technical advantages:

  • Finer segmentation: Developers can sample highly specific slices of the Vietnamese population. This includes generating precise cohorts, such as a retiree in Ho Chi Minh City, a smallholder farmer in the Mekong Delta, or a young engineer in Da Nang. The expanded dataset provides enough records in each category to train and test models reliably.
  • Larger workloads: The expanded scale supports bigger training runs, more comprehensive red-teaming, and population-level simulations that the original June 2026 release could not sustain on its own.
  • Reduced model collapse: Richer variation in narratives, skills, and behaviors lowers the risk of model collapse. This is a growing concern in AI pipelines where models trained on synthetic data can collapse onto a narrow set of templates if the training data lacks sufficient diversity.
  • Deeper coverage: While the earlier release sampled a set of major provinces and cities, the 1,000,000-persona version covers the entire country, enabling developers to build and test for users in every region.

The dataset remains fully synthetic, contains zero personally identifiable information, and is released openly under a permissive license on Hugging Face, making it compatible with NVIDIA NeMo libraries.

Sovereign AI and Vietnam's Digital Economy Goals

The expansion of this dataset aligns with Vietnam's broader digital transformation strategy. Under the newly approved Program for Digital Economy and Digital Society Development 2026-2030, the Vietnamese government targets the digital economy to reach approximately 30% of GDP by 2030.

Across industries such as banking, finance, healthcare, insurance, retail, and public services, organizations increasingly require realistic persona data for product testing, service personalization, simulation, and AI model evaluation. By expanding this dataset, FPT aims to build the data layer of Vietnam's sovereign AI—AI that is grounded in the country's own population, language, and values rather than imported from external sources.

Practical Applications in Agentic AI

As agentic AI emerges as a key frontier in technology—moving from simple prompt-and-response systems to autonomous agents that reason, act, and serve specific users—the demand for grounded, identity-aware data is growing. The expanded dataset provides the Vietnamese AI community with a foundation to build agents that understand their target users.

  • Customer Personas: Organizations can build customer-facing agents that draw on personas spanning various income brackets, regions, and financial-literacy levels. This allows agents to adjust their tone, product recommendations, and compliance handling for specific users, such as a retiree in Hanoi versus a young professional in Ho Chi Minh City.
  • Simulation and Evaluation: Developers can create simulation environments populated with realistic Vietnamese users and organizations. This enables red-teaming, bias-and-fairness evaluations across diverse demographic groups, and stress-testing of agent behavior across regions, ages, and occupations without using real customer data.
  • Domain-Specific AI Agents: The dataset enables specialized AI agents across regulated industries. In healthcare, agents can be evaluated against patient personas spanning different literacy levels, regions, and care settings. In public services, agents can be tested against citizen-personas across the country's administrative landscape.

The complete dataset is available for free download on Hugging Face.