FPT announced on 16 October 2026 (Hanoi) that it and NVIDIA have expanded Nemotron-Personas-Vietnam from 100,000 to 1 million records — a tenfold increase from the June 2026 initial release. The dataset is on Hugging Face.
The profiles are synthetic: FPT says they represent hypothetical Vietnamese individuals, are generated by AI, and are not collected from real people.
What each record holds
According to FPT, each profile has 21 data fields: six demographic attributes and fifteen contextual ones, covering age, gender, education, occupation, place of residence, skills and interests, plus descriptive passages about work, sports, travel and food.
Generation uses statistics from two Vietnamese official surveys — the 2024 Mid-Term Population and Housing Survey and the 2024 Vietnam Household Living Standards Survey (General Statistics Office / National Statistics Office) — so combinations of age, occupation, education and residence follow national population distributions. AI then adds personality traits, skills, interests and narratives.
FPT says it contributed source data, compute, research and local demographic/cultural expertise and helped evaluate quality; NVIDIA provided the methodology and tools for generating the synthetic profiles.
Coverage and Hugging Face licence
The June release focused on six provinces and cities. The expansion, FPT says, broadens coverage across more population groups, occupations and living circumstances in different regions of Vietnam. FPT also said that within four days of the first launch the dataset ranked among the Top 15 trending datasets on Hugging Face.
On licensing: FPT’s release only referred to “applicable license terms.” The Hugging Face dataset card lists the licence as CC BY 4.0 (cc-by-4.0 in the card metadata). The card’s dataset info reports 1,000,000 train examples, matching the million-record claim. State that licence because the card shows it.
Suggested uses — and the safety caveat
FPT frames the set as a way to train, test and evaluate AI apps with Vietnamese-like user diversity, including agentic multi-step assistants. Examples in the release (banking help for an elderly user; administrative form-filling) are suggested testing scenarios, not announced customer deployments.
FPT is explicit that synthetic profiles “do not replace the need to verify system accuracy and safety before deployment.” It also links the work to sovereign AI — more local data to research and evaluate systems for domestic users — especially in regulated sectors such as finance, healthcare, insurance and public services.
How the million-row card reads
The Hugging Face API/dataset_info block for nvidia/Nemotron-Personas-Vietnam lists a single train split with 1,000,000 examples and Vietnamese (vi) as the language tag, alongside synthetic/persona tags. That matches FPT’s public count. Treat download-size figures on the card as hosting metadata, not a quality score.
Context from earlier TSN coverage
NVIDIA’s Nemotron line and regional data partnerships are a recurring theme. TSN covered NVIDIA’s Nemotron 3.5 Super VL, NVIDIA and AI4Bharat’s Indic pretraining corpus, and Naver Cloud’s MoU with FPT on AI factory and GPU cloud. Today’s item is a synthetic persona dataset scale-up, not a new foundation model.
Why it matters
Open, licence-clear synthetic personas tuned to a country’s survey statistics are a practical answer when global chat logs poorly match local users — and when using real customer data for testing is legally or ethically hard. The tenfold jump mainly increases scenario coverage; it does not, by itself, prove models will behave safely in production.
What’s still unknown
- How demographic fidelity was audited beyond FPT’s internal evaluation
- Whether field schemas changed between the 100k and 1M releases
- Uptake outside Hugging Face downloads/trending metrics
What to watch
Watch the Hugging Face card for schema or licence changes, and any FPT follow-on on domain-specific evaluation suites built on the personas.
What this means
- 1 million synthetic Vietnamese personas, up from 100,000 (tenfold).
- Not real people; survey-conditioned AI generation (FPT).
- Licence on Hugging Face: CC BY 4.0.
Sources
- FPT, “FPT and NVIDIA Expand Synthetic User Dataset Tenfold to Advance AI Development”, 16 October 2026
- Hugging Face, nvidia/Nemotron-Personas-Vietnam (dataset card; licence CC BY 4.0)
Related stories
- Naver Cloud and FPT Sign an MOU to Grow AI Factory and GPU Cloud Business Across Asia-Pacific
- NVIDIA Releases Nemotron 3.5 Super VL, an Open 120B Multimodal MoE for Agents
- NVIDIA and AI4Bharat Detail a 2.8-Trillion-Token Indic Pretraining Corpus
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
