AI Wiki

Explore AI Wiki

Citation and evidence

Nemotron-Personas

16 min full readUpdated 30 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision history

Topics

Data & DatasetsLarge Language ModelsNVIDIATraining & Optimization

Cite this article

Nemotron-Personas is a collection of open synthetic persona datasets published by NVIDIA on Hugging Face. Each dataset describes a population of fictional adults whose ages, locations, education, occupations, household situations and other attributes are sampled to match a country's official statistics, and whose written profiles (professional, sports, arts, travel, culinary and similar "personas") are generated by a large language model. NVIDIA positions the collection as seed material for synthetic data generation and for sovereign AI work, where developers want training data that reflects a specific country's people rather than a generic, English-centric user.[2][3] The first release, a United States dataset, appeared in June 2025.[3] As of 28 October 2026 the Hugging Face collection lists twelve country datasets: the USA, Japan, India, Singapore, Brazil, France, South Korea, El Salvador, Vietnam, Belgium, Germany and the United Kingdom, all under the CC BY 4.0 license.[2] Several were built with local partners, including Pleias, KU Leuven, WideLabs, FPT, Locai Labs and the German Soofi project.[9][10][12][13][14][15][16]

The datasets are part of NVIDIA's broader Nemotron program of open models and data. NVIDIA has used them inside its own model training, for example to seed tool-use conversations for Nemotron 3 Super and Japanese tool-calling data for Nemotron-Nano-9B-v2-Japanese.[18][19] In October 2026 NVIDIA published a tutorial showing how the France persona set can be used to fine-tune Nemotron 3.5 Lightning into a French public-service assistant.[1]

Background

Persona-conditioned generation is a common technique for making synthetic training data more varied. Instead of asking a model for thousands of user questions with no context, which tends to produce many near-duplicates, each request is conditioned on a different described person. A widely cited earlier example is Persona Hub, introduced in a June 2024 paper by Tao Ge and colleagues, which curated 1 billion personas automatically from web data and used them to synthesize math problems, instructions, knowledge texts, game characters and tool definitions.[20]

NVIDIA's approach differs in where the personas come from. Rather than extracting personas from web text or asking an LLM to invent them, Nemotron-Personas first samples structured demographic records from statistical models fitted to census and administrative data, and only then asks an LLM to write the narrative fields.[3][17] NVIDIA's dataset cards argue that this makes the personas more representative along axes such as age (older people), geography (rural residents), education and occupation than earlier persona datasets, and that it helps reduce bias and model collapse when synthetic data is reused for training.[3] Model collapse, the degradation that occurs when models are trained recursively on generated data, was described in a 2024 Nature paper by Ilia Shumailov and colleagues.[21] The USA card contrasts the dataset's census-shaped age pyramid with the "bell curve distribution typically produced by an LLM alone."[3]

How the datasets are built

Every dataset in the collection is produced with NeMo Data Designer, NVIDIA's synthetic data generation system, which is open source on GitHub under the Apache 2.0 license.[17][24] An NVIDIA developer note describes a four-stage pipeline that is shared across locales:[17]

StageWhat it producesNotes from NVIDIA's description
1. Personality samplingBig Five (OCEAN) trait scoresFive T-scores are sampled (mean 50, standard deviation 10, clipped to 20-80), bucketed into labels from "very low" to "very high", and mapped to prose descriptions
2. Demographic samplingAge, sex, marital status, education, occupation, location and locale-specific fieldsDrawn from probabilistic graphical models (PGMs) fitted to census tables, administrative records and surveys, so that attributes co-vary realistically
3. Persona attributesCultural background, skills and expertise, career goals, hobbies and interestsOne structured-output LLM call, validated against a Pydantic schema
4. Persona descriptionsShort narrative personas (professional, sports, arts, travel, culinary and others)A second structured-output LLM call; the system prompt tells the model not to mention cultural heritage directly

Expanded article table

The PGM engine behind stage 2 was later released as an open-source library, NeMo SDG-PGMs, whose GitHub repository was created in April 2026.[17][25] The LLM used for the narrative stages varies by release. The original USA version used Mixtral-8x22B, which the v1.1 update replaced with gpt-oss-120b.[3] The Japan, India, Singapore, Brazil and El Salvador releases also used gpt-oss-120b; France used an NVFP4 version of NVIDIA Nemotron 3 Super; Korea, Belgium, the UK and Germany used Google's Gemma 4 31B instruction-tuned model; and Vietnam used FPT's SaoLa-Small-32B.[5][7][8][9][10][11][12][13][14][15][16]

The public releases leave out some fields that the generator produces. The USA card says names and synthetic addresses are excluded even though names are woven into the persona text, and that finance and healthcare personas, described as being of interest to enterprise customers, are also withheld.[3] Most cards state that an extended version of the dataset is, or will soon be, available inside NeMo Data Designer, and NVIDIA's developer note shows the personas being loaded through Data Designer's PersonSampler from NVIDIA's NGC catalog.[3][17] All the datasets cover adults only.[3][15]

Datasets in the collection

The table lists the twelve datasets in the Hugging Face collection as of 28 October 2026, with figures taken from each dataset card. "Personas" counts individual persona descriptions (several per record).

DatasetCard release dateDeveloper(s)RecordsPersonasLanguage(s)Statistical groundingNarrative LLM
Nemotron-Personas-USA9 June 2025 (v1.1: 28 Oct 2025)NVIDIA1M (100k in v1.0)6MEnglishUS Census Bureau American Community Surveygpt-oss-120b (v1.0: Mixtral-8x22B) [3]
Nemotron-Personas-Japan23 Sept 2025NVIDIA1M6MJapaneseStatistics Bureau of Japan census datagpt-oss-120b [5]
Nemotron-Personas-India13 Oct 2025NVIDIA3M21MEnglish, Hindi (Devanagari), Hindi (Latin script)2011 Census of Indiagpt-oss-120b [7]
Nemotron-Personas-Brazil26 Jan 2026WideLabs and NVIDIA1M6MPortugueseIBGE census and labor datagpt-oss-120b [9]
Nemotron-Personas-Singapore27 Jan 2026NVIDIA148k888kEnglishSingapore Department of Statisticsgpt-oss-120b [8]
Nemotron-Personas-France15 Mar 2026Pleias and NVIDIA1M6MFrenchINSEE census and name filesNemotron 3 Super (NVFP4) [10]
Nemotron-Personas-Korea20 Apr 2026NVIDIA (NAVER Cloud contributed seed data)1M7MKoreanKOSIS, Supreme Court, NHIS and KREI dataGemma 4 31B-it [11]
Nemotron-Personas-El-Salvador3 June 2026WideLabs and NVIDIA148kabout 1MSalvadoran Spanish2024 population and housing census (BCR/ONEC)gpt-oss-120b [12]
Nemotron-Personas-Vietnam4 June 2026 (v2.0: 15 Oct 2026)NVIDIA with FPT Smart Cloud and FPT's Quantum AI and Cyber Security Institute1M (v2.0)6M (v2.0)VietnameseGeneral Statistics Office dataSaoLa-Small-32B [13]
Nemotron-Personas-Belgium16 June 2026Pleias, NVIDIA and KU Leuven1.2M (300k per language)1.8M, each in four languagesDutch, French, German, EnglishStatbel Census 2021Gemma 4 31B-it [14]
Nemotron-Personas-UK20 Oct 2026Locai Labs and NVIDIA1M6MEnglishCensus 2021 (England and Wales; Northern Ireland) and Scotland's Census 2022Gemma 4 31B-it [15]
Nemotron-Personas-Germany20 Oct 2026Soofi and NVIDIA1M6MGermanZensus 2022Gemma 4 31B-it [16]

Expanded article table

The October 2026 NVIDIA tutorial describes the program as including "locale-grounded persona datasets for 12 countries," which matches the twelve entries in the collection.[1][2]

Record structure

The USA dataset illustrates the basic schema: 22 fields per record, made up of six persona fields (professional_persona, sports_persona, arts_persona, travel_persona, culinary_persona and a summary persona) and 16 contextual fields such as cultural_background, skills_and_expertise, hobbies_and_interests, career_goals_and_ambitions, sex, age, marital_status, education_level, bachelors_field, occupation, city, state, zipcode and country, plus a unique identifier.[3] The card reports about 936 million tokens across the 1 million records and more than 560 occupations.[3]

Later releases add locale-specific fields and persona types. India has seven persona fields per record, including a "linguistic" persona, delivered in three separate language and script splits.[7] Korea adds a "family" persona for seven persona fields per record, and its contextual fields include military service status, family type and housing type.[11][23] The UK and Germany releases separate six "persona-attribute" fields from the contextual fields, for 23 and 27 fields respectively.[15][16]

Notable releases

United States. The first dataset was published on 9 June 2025 under the name nvidia/Nemotron-Personas, with 100,000 records. Version 1.1, released on 28 October 2025, expanded it to 1 million records (about 0.94 billion tokens), switched the narrative model to gpt-oss-120b and renamed it Nemotron-Personas-USA to distinguish it from the regional datasets.[3] NVIDIA announced the expanded version in a Hugging Face blog post the same day, describing 6 million personas spanning all 50 states and territories.[4] The old repository name now redirects to the USA dataset on Hugging Face. The card's citation lists Yev Meyer and Dane Corneil as authors.[3] The card acknowledges simplifying assumptions, for example that occupation is independent of ZIP code given education, age and sex, and notes that the Census Bureau does not publish comprehensive statistics on gender separately from sex.[3]

Japan. Released on 23 September 2025, Nemotron-Personas-Japan was the first non-US dataset. NVIDIA's announcement described it as the start of "a global collection of synthetic persona datasets and playbooks" for sovereign AI.[6] It covers all 47 prefectures, and NVIDIA says it extended the official statistics with finer education categories, extra occupation categories, life-stage information and varying levels of digital literacy by age.[5][6]

India. The India release (13 October 2025) is grounded in the 2011 census. Because Telangana was created in 2014, after that census, the authors rebuilt its statistics from the districts it was carved out of. The dataset covers all 36 states and union territories and 640 districts, with about 7.7 billion tokens in total.[7]

Vietnam. FPT announced the Vietnamese dataset on 3 June 2026, describing it as built "in collaboration with NVIDIA" and grounded in Vietnam's 2024 statistical yearbook and its post-consolidation provincial map; FPT's announcement cited 900,000 synthetic personas for the initial release.[22] Version 2.0 of the dataset, dated 15 October 2026 on the card, contains 1 million records and covers all 34 provinces and centrally administered cities that exist after the 2025 administrative merger.[13]

Belgium. The Belgian dataset is the most multilingual entry. Each of its 300,000 personas is delivered in four splits (nl_BE, fr_BE, de_BE and en_BE) rather than tagged with a language field. Each persona's primary language is set from the official language of its commune, the language probabilities of its name and its cultural heritage, which the card says yields 58% Dutch-speaking, 39% French-speaking and 1% German-speaking personas, with 2% speaking another home language. The card warns that the non-English editions are machine-translated and that religion is not modeled.[14]

United Kingdom. Nemotron-Personas-UK combines three censuses, each with its own PGM, and allocates records in proportion to each nation's adult population: 889,000 records for England and Wales, 83,000 for Scotland and 28,000 for Northern Ireland, across 372 local authorities.[15] It was built with Locai Labs, a London company that sells an on-premises AI computer called the Locai 1; the card says generation ran on 18 NVIDIA GPUs across cloud and local compute, including a Locai 1.[15][30]

Germany. The German dataset, released alongside the UK one on 20 October 2026, was developed with Soofi (Sovereign Open Source Foundation Models), a German and European project building open foundation models that is coordinated by the German AI Association and funded by the Federal Ministry for Economic Affairs and Energy.[16][29] It covers all 400 districts across the 16 federal states and samples occupations from the 2010 German occupation classification (KldB 2010).[16]

Use in Nemotron model training

NVIDIA's own models use the personas as seed data. The Nemotron 3 Super technical report says the team built long-context training samples "by concatenating records from" Nemotron-Personas-USA, seeded a simulated user in its tool-calling trajectory pipeline with "a persona sampled from Nemotron-Personas-USA," and added random personas to vary the scenarios in its synthetic formal-logic data.[18] A Data Designer developer note adds that a sensitive-category safety refusal dataset in NVIDIA's safety blend uses the personas as seeds, and that instruction-following and general-chat data for Nemotron Nano v3 and Super v3 follow the same persona-seeded recipe.[17] (See Nemotron 3 for those models.)

The model card for NVIDIA-Nemotron-Nano-9B-v2-Japanese says the model was further trained from Nemotron-Nano-9B-v2 using Japanese tool-calling data created with Nemotron-Personas-Japan, and lists a 4-billion-token synthetic Japanese tool-calling set that uses the personas as one of its seed datasets.[19] NVIDIA's developer note claims the model ranks first on the Nejumi LLM Leaderboard; that ranking is NVIDIA's own statement.[17]

Local-rules fine-tuning workflow

On 27 October 2026 NVIDIA published a technical blog post by Sophia Abbassi and Shashank Verma, "Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languages," promoted the same day by the NVIDIA AI account on X.[1][28] The post walks through building a French passport and vehicle-registration support agent from Nemotron-Personas-France. Its motivating example is a user who cannot travel for health reasons. According to NVIDIA, a model without post-training told the user to send the application by registered post, which conflicts with official guidance that a government official must make a home visit; the fine-tuned model instead described requesting a home visit with a medical certificate less than three months old.[1] The post also notes that the right answer can depend on where the persona lives: the passport tax stamp costs 86 euros in mainland France but 43 euros in French Guiana, which matches the French government's service-public.gouv.fr page.[1][27]

The workflow is published as the french-public-service-assistant example in the NVIDIA-NeMo/Nemotron GitHub repository (on its main branch), and NVIDIA says it targets a single NVIDIA DGX Station GB300 or DGX Spark, adaptable to other NVIDIA GPU systems.[1][26]

StepWhat happens
1. Behavior contract and holdoutDefine audience, task, language, grounding, safety rules and output format; NVIDIA recommends 100 to 300 human-written or de-identified test cases with expected behavior
2. Synthetic conversationsSample 2,000 personas across departments and occupations; a Data Designer step turns each into a customer message and an agent reply, with Nemotron 3.5 Lightning as the generator (served locally with vLLM in the DGX Station setup) and a separate judge model checking answers; facts come only from paired official procedure text
3. FilteringDrop rows with formatting problems, leftover reasoning or leaked persona names, then split 90/10 by persona ID so no persona appears in both sets
4. Fine-tuningLoRA training with a validated NeMo Automodel recipe for Lightning
5. Merge, serve and compareMerge the adapter into the base checkpoint, serve base and tuned models side by side, and compare their answers on the frozen prompts, preferably with human reviewers

Expanded article table

The post stresses a configuration detail: Nemotron 3.5 Lightning produces reasoning traces by default, and thinking must be disabled in the provider's chat-template settings or reasoning text ends up in the training data.[1] It tells developers to use persona fields only to vary how simulated customers communicate, not to infer their needs or competence from sensitive attributes, and to exclude names and unnecessary demographic details from the final training data.[1] The repository also ships agent skills meant to help coding agents assemble these steps into a workflow.[1]

Pleias and Lamina

The Paris-based lab Pleias co-developed the France dataset and, with KU Leuven, the Belgium dataset.[10][14] According to NVIDIA's October 2026 post, Pleias uses three Nemotron-Personas datasets, including France and Belgium, in its work on Lamina, a spine-care assistant made by a Pleias partner. The personas turn a small number of real patient cases into varied training conversations and let the assistant be tested across patient profiles. NVIDIA quotes Pleias as saying, "Better grounded data is what makes a small efficient regional-specific model viable."[1] The post adds that persona grounding is meant to widen coverage and reduce blind spots in training data and is "not a guarantee of fairness or proper representation."[1]

Limitations

The dataset cards list several caveats. The statistical models depend on what each statistics office publishes, so the authors make independence assumptions where joint distributions are unavailable.[3][14] Coverage is limited to adults.[3][15] Names, addresses and some sensitive attributes are excluded from the public releases.[3][14] The narrative fields are LLM-generated, so their quality depends on the model used, and the Belgian non-English editions are machine-translated.[14] The census vintages also differ widely between countries, from India's 2011 census to the 2025 French census used for the France dataset.[7][10] Persona text is synthetic, and each card states that any resemblance to real people is coincidental.[3][14]

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Sophia Abbassi and Shashank Verma, "Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languages," NVIDIA Technical Blog, 27 October 2026. developer.nvidia.com/...s-not-just-local-languages
  2. ^1 ^2 ^3NVIDIA, "Nemotron-Personas" collection, Hugging Face (accessed 28 October 2026). huggingface.co/...nemotron-personas
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19NVIDIA, "Nemotron-Personas-USA" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-USA
  4. ^Will Jennings, Dane Corneil and Yev Meyer, "Nemotron-Personas-USA: Synthesized Data for Sovereign AI," Hugging Face blog, 28 October 2025. huggingface.co/...nemotron-personas-usa
  5. ^1 ^2 ^3NVIDIA, "Nemotron-Personas-Japan" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Japan
  6. ^1 ^2Atsunori Fujita et al., "Nemotron-Personas-Japan: Synthesized Data for Sovereign AI," Hugging Face blog, 23 September 2025. huggingface.co/...nemotron-personas-japan
  7. ^1 ^2 ^3 ^4 ^5NVIDIA, "Nemotron-Personas-India" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-India
  8. ^1 ^2NVIDIA, "Nemotron-Personas-Singapore" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Singapore
  9. ^1 ^2 ^3WideLabs and NVIDIA, "Nemotron-Personas-Brazil" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Brazil
  10. ^1 ^2 ^3 ^4 ^5Pleias and NVIDIA, "Nemotron-Personas-France" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-France
  11. ^1 ^2 ^3NVIDIA, "Nemotron-Personas-Korea" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Korea
  12. ^1 ^2 ^3WideLabs and NVIDIA, "Nemotron-Personas-El-Salvador" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-El-Salvador
  13. ^1 ^2 ^3 ^4NVIDIA, FPT Smart Cloud and FPT Quantum AI & Cyber Security Institute, "Nemotron-Personas-Vietnam" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Vietnam
  14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Pleias, NVIDIA and KU Leuven, "Nemotron-Personas-Belgium" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Belgium
  15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Locai Labs and NVIDIA, "Nemotron-Personas-UK" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-UK
  16. ^1 ^2 ^3 ^4 ^5 ^6Soofi and NVIDIA, "Nemotron-Personas-Germany" dataset card, Hugging Face. huggingface.co/...Nemotron-Personas-Germany
  17. ^1 ^2 ^3 ^4 ^5 ^6 ^7NVIDIA, "Designing Nemotron-Personas: Multi-Locale Synthetic Personas Powering Nemotron Training," NeMo Data Designer documentation. docs.nvidia.com/...designing-nemotron-personas
  18. ^1 ^2NVIDIA, "Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning," technical report. research.nvidia.com/...-Super-Technical-Report.pdf
  19. ^1 ^2NVIDIA, "NVIDIA-Nemotron-Nano-9B-v2-Japanese" model card, Hugging Face. huggingface.co/...DIA-Nemotron-Nano-9B-v2-Japanese
  20. ^Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi and Dong Yu, "Scaling Synthetic Data Creation with 1,000,000,000 Personas," arXiv:2406.20094, 28 June 2024. arxiv.org/...2406.20094
  21. ^Ilia Shumailov et al., "AI models collapse when trained on recursively generated data," *Nature* 631, 755-759 (2024). nature.com/...s41586-024-07566-y
  22. ^FPT AI Factory, "Introducing Nemotron Personas Vietnam Dataset: Sovereign AI Grounded in Vietnamese Reality," 3 June 2026. factory.fpt.ai
  23. ^PyTorchKR, "Nemotron-Personas-Korea" introduction (in Korean), 27 April 2026. discuss.pytorch.kr/...9908
  24. ^NVIDIA-NeMo, "DataDesigner" repository, GitHub. github.com/...DataDesigner
  25. ^NVIDIA-NeMo, "SDG-PGMs" repository, GitHub. github.com/...SDG-PGMs
  26. ^NVIDIA-NeMo, "Nemotron" repository, use-case example "french-public-service-assistant," GitHub. github.com/...french-public-service-assistant
  27. ^Service-Public.fr, "Passeport" (fees and procedure for adults in France), French government. service-public.gouv.fr/...F14929
  28. ^NVIDIA AI (@NVIDIAAI), post on X, 27 October 2026. x.com/...2115116486671176161
  29. ^Fraunhofer IIS, "Soofi: Sovereign AI from Europe." iis.fraunhofer.de/...soofi
  30. ^Locai Labs, company website. locailabs.com

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision · v1 · 3,178 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Each of the twelve datasets was checked against its Hugging Face dataset card, and the fine-tuning workflow against NVIDIA's October 2026 technical blog, the companion GitHub example and the Persona Hub paper, October 2026.

Cite this page: AI Wiki. "Nemotron-Personas." aiwiki.ai, updated 28 Oct 2026, fact-checked 28 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/nemotron_personas

Suggest edit

Where to go next

Topics:Data & DatasetsLarge Language ModelsNVIDIATraining & Optimization