FPT and NVIDIA Expand Nemotron-Personas-Vietnam Tenfold to One Million Synthetic Profiles

FPT and NVIDIA have expanded Nemotron-Personas-Vietnam, an open dataset of synthetic Vietnamese user profiles, from 100,000 to 1 million records, FPT announced in Hanoi on October 16. The profiles describe fictional people, none collected from real individuals, and are meant to give researchers and developers more material to train, test and stress AI applications built for Vietnamese users.

Each profile carries 21 fields: six portrait fields and 15 contextual ones covering age, gender, education, occupation, place of residence, skills and hobbies, plus short narrative passages about work, sport, travel and food. The demographic backbone comes from two 2024 surveys run by the former General Statistics Office, now the Statistics Department under the Ministry of Finance: the mid-term population and housing census and the household living standards survey. AI then adds personality traits, skills and descriptive text.

FPT supplied source data, computing infrastructure, research capacity and knowledge of Vietnamese society and culture, and helped check quality; NVIDIA provided the methods and tools for generating the personas. The first release in June covered six provinces and cities and entered the top 15 trending datasets on Hugging Face within four days. The new version broadens coverage across regions, occupations and living situations.

What it means in practice

Synthetic personas let teams rehearse realistic scenarios. FPT’s example is a bank assistant guiding an elderly customer through checking transaction history on a phone, a test that reduces the need to use real customer records. They can also drive multi-step agent tests in banking, health, insurance and public services, though FPT stresses they do not replace accuracy and safety checks before deployment. The release builds on local work such as Viettel’s Nemotron fine-tuning for Vietnamese legal agents and Hanoi’s Agent City initiative for public-service agents.

For regulators, the dataset arrives as the generative-AI sandbox sets out what testing does and does not permit. Persona-based test suites could become a practical way to show supervisors that an agent behaves consistently across age groups, regions and income levels before it reaches the public.

What to watch: licence terms for commercial use, independent checks of how closely the profiles mirror real population distributions, and whether banks and ministries publish results from persona-based testing.