Vietnamese technology company FPT has expanded Nemotron-Personas-Vietnam, its open synthetic-data dataset developed using NVIDIA’s Nemotron-Personas methodology, to 1 million personas. Grounded in Vietnam’s official demographic and labor statistics, the dataset is designed to help developers build, train, and evaluate AI models and agents that better reflect the country’s population, regional differences, occupations, and cultural context. Available through Hugging Face, the resource offers a synthetic alternative to real customer records for applications such as agent simulation, product testing, and AI evaluation, while raising broader questions about representativeness and the reliability of synthetic training data.
FPT Expands Vietnamese Synthetic AI Dataset to One Million Personas
As companies deploy AI agents across customer service, financial services, healthcare, and public-sector applications, developers face a persistent challenge: building systems that understand the people they are intended to serve. Models trained predominantly on generalized or foreign-language data may fail to capture local terminology, regional differences, cultural expectations, and socioeconomic context.
FPT is addressing that gap by expanding Nemotron-Personas-Vietnam, an open dataset of synthetic Vietnamese personas, to 1 million records. First released in June 2026 in collaboration with NVIDIA, the dataset uses NVIDIA’s Nemotron-Personas methodology to generate synthetic profiles grounded in official Vietnamese demographic and labor statistics.
The expansion is intended to give developers, researchers, and enterprises a larger resource for training, evaluating, and simulating AI systems for Vietnamese users. It also reflects a wider effort to develop AI infrastructure and datasets that are more closely aligned with local populations and languages.
Building AI personas from official statistics
According to FPT, Nemotron-Personas-Vietnam draws on data from Vietnam’s General Statistics Office, including the 2024 Population and Housing Census and the 2024 Vietnam Household Living Standards Survey. The project also incorporates contributions from FPT Smart Cloud and FPT Corporation’s Quantum AI & Cyber Security Institute.
The dataset contains 21 fields: six persona fields and 15 contextual fields covering demographic and other attributes. It includes 13,000 unique full-name combinations and 20 occupational categories derived from labor survey data.
Profiles are designed to represent variations in age, region, occupation, income, education, and cultural context. Developers can use these attributes to construct test scenarios for different population groups—for example, evaluating whether an application works appropriately for users with different levels of financial literacy or contrasting regional needs.
The personas are synthetic rather than records of identified individuals. FPT says the dataset contains no personally identifiable information and is released openly under a permissive license through Hugging Face, with compatibility for NVIDIA NeMo libraries.
Synthetic data can reduce reliance on real customer records during development and testing. However, the absence of personal information does not by itself guarantee that generated profiles accurately represent every community or avoid statistical bias. Those outcomes depend on source-data quality, generation methods, validation, and how developers use the dataset.
Why scaling to one million matters
The expansion increases the number of profiles available for sampling, testing, and simulation. FPT says the updated dataset extends coverage across the entire country, compared with the earlier release’s selection of major provinces and cities.
Broader coverage could help developers test applications across a wider range of geographic and socioeconomic scenarios. A financial-services provider, for instance, could simulate interactions involving customers with different incomes, occupations, or levels of financial knowledge. Public-service teams could test whether digital workflows remain understandable across varied user profiles.
The larger dataset may also support more extensive model evaluations and red-team exercises. Developers can examine how agents respond to different user characteristics, whether outputs vary unfairly across demographic groups, and how systems handle edge cases before deployment.
FPT also argues that greater narrative and behavioral diversity can help reduce repetitive patterns in synthetic training data. That is a potential benefit rather than a guaranteed result: increasing dataset size alone does not eliminate model collapse or ensure diversity. The quality and independence of generated examples, along with the training process, remain important.
Supporting Vietnam-focused and sovereign AI
FPT positions the expansion as part of a broader effort to strengthen Vietnam’s AI capabilities using data grounded in the country’s own population and social context. This approach aligns with the wider concept of sovereign AI, in which countries and organizations seek greater control over the data, infrastructure, and systems used to build AI applications.
The company points to Vietnam’s digital-economy ambitions and the need for locally relevant technology across banking, healthcare, insurance, retail, and public services. Synthetic personas could help organizations prototype services, evaluate interfaces, and test agent behavior without directly exposing real customer information.
FPT’s cited policy context includes Vietnam’s 2026–2030 digital economy and digital society development program, which targets the digital economy reaching approximately 30% of GDP by 2030. That figure is a government target referenced in the supplied announcement, not a forecast of a guaranteed outcome.
A foundation for agent simulation and evaluation
As AI agents move beyond answering prompts to performing multistep tasks, testing must account for how systems respond to different users and circumstances. A synthetic persona dataset can supply repeatable scenarios for testing conversational behavior, product recommendations, service workflows, and potential demographic disparities.
In regulated industries, developers could use synthetic profiles to evaluate whether agents explain financial products appropriately, handle healthcare-related interactions consistently, or guide citizens through public-service processes. Such tests can reveal weaknesses before systems encounter real users, although synthetic simulations cannot replace real-world validation, expert review, or compliance testing.
The dataset’s long-term value will depend on transparent documentation, representative sampling, meaningful evaluation, and continued maintenance. Expanding to one million personas gives developers a larger starting point, but not proof that every profile or resulting AI system is representative.
For Vietnam’s AI ecosystem, Nemotron-Personas-Vietnam offers a shared, openly accessible resource for developing and testing locally relevant AI. Its significance will ultimately be measured by whether it helps teams build systems that perform more reliably across the country’s diverse population.
Market Landscape
Synthetic data is becoming an important component of AI development, particularly where real customer information is difficult to access, sensitive, or costly to label. Persona datasets can support model evaluation, application testing, agent simulation, and bias assessment without directly using identifiable customer records.
FPT’s expansion targets a specific gap: locally grounded synthetic data for Vietnamese users. By using official demographic and labor statistics, the project aims to make AI development more relevant to regional, occupational, and socioeconomic differences within Vietnam.
The broader opportunity lies in combining synthetic data with enterprise AI platforms, evaluation frameworks, and agent-development tools. However, dataset scale alone does not establish representativeness. Developers will need to examine coverage, sampling bias, cultural accuracy, and performance against real-world benchmarks before relying on synthetic personas for consequential decisions.
Top Insights
- FPT has expanded Nemotron-Personas-Vietnam to one million synthetic personas designed to support AI development for Vietnamese users.
- The dataset contains 21 fields, 13,000 unique full-name combinations, and 20 occupational categories based on labor survey data.
- Its source material includes Vietnam’s 2024 Population and Housing Census and Vietnam Household Living Standards Survey.
- The open dataset is available on Hugging Face and is compatible with NVIDIA NeMo libraries, according to FPT.
- The expanded resource could support agent simulation and fairness testing, but dataset size alone cannot guarantee representativeness or eliminate synthetic-data bias.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
