![\"Nemotron-Personas-India\"](\"https://cdn-uploads.huggingface.co/production/uploads/67ee906af503c8811bebcbf3/VXDHXVACuC0FOkOgkJtZI.png\")
India represents one of the world's largest AI opportunities — with over 700 million internet users, a multitude of languages, and a rapidly growing developer ecosystem. Yet, most open datasets reflect Western norms and English-only contexts, creating a data gap that limits AI adoption in India's multilingual, multi-script environment.
\nToday, we're releasing Nemotron-Personas-India, the first open synthetic dataset of Indic personas aligned to India's real-world demographic, geographic, and cultural distributions. Licensed under CC BY 4.0, this dataset offers a privacy-preserving, regulation-ready foundation for scaling AI systems that reflect Indian society—without relying on sensitive personal data.
\nBuilt with NeMo Data Designer, NVIDIA's enterprise-grade synthetic data generation microservice, Nemotron-Personas-India extends our global collection of Sovereign AI datasets. It builds on the success of our US and Japan persona datasets and includes new features designed specifically for India's culturally rich landscape.
\nThis dataset integrates seamlessly with Nemotron models and other open-source LLMs, making it easy to fine-tune AI systems for Indian use cases—from multilingual chatbots to culturally-grounded specialized copilots.
\nThis release complements our earlier suite of Hindi evaluation datasets — including ChatRAG-Hi, IFEval-Hi, MT-Bench-Hi, GSM8K-Hi, and BFCL-Hi — supporting a complete pipeline from synthetic data generation to rigorous model evaluation for Indian AI systems.
\n![\"image/png\"](\"https://cdn-uploads.huggingface.co/production/uploads/67ee906af503c8811bebcbf3/-ef4HnfGJO5m4ov3QXcL5.png\")
Produced using NeMo Data Designer, NVIDIA's microservice for synthetic data generation. This compound AI system enables generation with complex Jinja templating, Pydantic validation, structured outputs, automated retries, and supports multiple generation backends – the necessary tooling to scale a synthetic dataset of this size. We also leveraged the following models:
\nThis dataset was aligned to India’s official demographic distributions from the 2011 Census and expanded to include attributes essential for trustworthy AI training:
\nNo real names. No re-identification risk.
\nAll personas are fully synthetic. While grounded in real-world distributions from the 2011 Census and Parsed Indian Electoral Rolls data, no data is tied to any living or deceased individual. This ensures developers can safely train AI systems without privacy risks or regulatory barriers.
\nBuilt for India, Ready for the World
\nNemotron‑Personas‑India is designed for developers building Sovereign AI systems for the Indian market, as well as global teams looking to adapt models to India’s unique linguistic, cultural, and social context.
\nMost open datasets today reflect English-speaking, Western norms—limiting AI performance in India’s multilingual, multi-script, and demographically complex environments.
\nWith Nemotron‑Personas‑India, teams can:
\nIndia's 1.4 billion people speak hundreds of languages and live across vast cultural, economic, and geographic divides. India's National AI Portal estimates over 7,000 AI startups and research institutions are working to build locally relevant AI systems, and the Digital India initiative and government programs like IndiaAI are accelerating adoption.
\nBut progress is constrained by a fundamental gap: high-quality, culturally grounded training data that reflects India's demographic reality. Without representative datasets, AI systems struggle with code-switching between English and Hindi, fail to understand regional occupational categories, and miss cultural context essential for trust and adoption.
\nThe dataset improves diversity of synthetically-generated data, mitigates biases, and prevents model collapse (degradation caused by uncurated training on another model's outputs) by reflecting India's real geographic and demographic distributions.
\nNemotron-Personas-India supports Indian model builders in developing Sovereign AI systems that incorporate important region-specific demographics and cultural context.
\nWant to build AI systems that understand India's culture, languages, and people?
\nTo start experimenting today:
\n\n\n# English personas\n = \n# Hindi personas in Devanagari\n = \n# Hindi personas in Latin\n = \nWhether you're an Indian model builder developing Sovereign AI or a global developer seeking better regional adoption, Nemotron-Personas-India provides the authentic, privacy-safe foundation your applications need.
\nDownload it. Fine-tune it. Build AI that understands India. If you’re ready to go deeper, an extended version of Nemotron-Personas-India (which includes e.g., first/last names, religion, and synthetic addresses) is available in NeMo Data Designer.
\n", "content_markdown": " *A compound AI approach to Indian personas grounded in real-world distributions*\n\n## Open Data for India's AI Future\n\nIndia represents one of the world's largest AI opportunities — with over 700 million internet users, a multitude of languages, and a rapidly growing developer ecosystem. Yet, most open datasets reflect Western norms and English-only contexts, creating a data gap that limits AI adoption in India's multilingual, multi-script environment.\n\nToday, we're releasing [Nemotron-Personas-India](https://huggingface.co/datasets/nvidia/Nemotron-Personas-India), the first open synthetic dataset of Indic personas aligned to India's real-world demographic, geographic, and cultural distributions. Licensed under **CC BY 4.0**, this dataset offers a privacy-preserving, regulation-ready foundation for scaling AI systems that reflect Indian society—without relying on sensitive personal data.\n\nBuilt with [NeMo Data Designer](https://docs.nvidia.com/nemo/microservices/latest/generate-synthetic-data/index.html), NVIDIA's enterprise-grade synthetic data generation microservice, Nemotron-Personas-India extends our global collection of [Sovereign AI](https://www.nvidia.com/en-us/lp/industries/global-public-sector/sovereign-ai-technical-overview/) datasets. It builds on the success of our [US](https://huggingface.co/datasets/nvidia/Nemotron-Personas) and [Japan](https://huggingface.co/datasets/nvidia/Nemotron-Personas-Japan) persona datasets and includes new features designed specifically for India's culturally rich landscape.\n\nThis dataset integrates seamlessly with **Nemotron models** and other open-source LLMs, making it easy to fine-tune AI systems for Indian use cases—from multilingual chatbots to culturally-grounded specialized copilots.\n\nThis release complements our earlier suite of Hindi evaluation datasets — including [ChatRAG-Hi](https://huggingface.co/datasets/nvidia/ChatRAG-Hi), [IFEval-Hi](https://huggingface.co/datasets/nvidia/IFEval-Hi), [MT-Bench-Hi](https://huggingface.co/datasets/nvidia/MT-Bench-Hi), [GSM8K-Hi](https://huggingface.co/datasets/nvidia/GSM8K-Hi), and [BFCL-Hi](https://huggingface.co/datasets/nvidia/BFCL-Hi) — supporting a complete pipeline from synthetic data generation to rigorous model evaluation for Indian AI systems.\n\n## What’s in the Dataset?\n\n[](https://cdn-uploads.huggingface.co/production/uploads/67ee906af503c8811bebcbf3/-ef4HnfGJO5m4ov3QXcL5.png)\n\n- **21 million personas total** (3M records × 7 personas each)\n- **Multilingual support:** English and Hindi, in both Devanagari and Latin scripts\n- **27 fields per record:** Persona traits + contextual attributes grounded in official census and labor statistics, including age, gender, education, occupation, state, district, and more\n- **7.7 billion tokens total**, including 2.9B persona tokens\n - English: 1B tokens total, 394M persona tokens\n - Hindi (Devanagari): 4.7B tokens total, 1.8B persona tokens\n - Hindi (Latin): 2B tokens total, 746M persona tokens\n- **\\~560k unique full names**, reflecting India's vast linguistic diversity\n- **2.9k occupational categories**, including informal, formal, and traditional sectors\n- **All 36 states of India** and 640 districts represented\n- **Natural language fields** cultural background, linguistic background, skills and expertise, hobbies and interests\n- **Persona types:** Includes general, professional, linguistic, culinary, sports, arts, and travel personas\n- **Licensed under CC BY 4.0** for commercial and non-commercial use\n\n## How We Built It\n\n### Data Generation Pipeline\n\nProduced using [NeMo Data Designer](https://docs.nvidia.com/nemo/microservices/latest/generate-synthetic-data/index.html), NVIDIA's microservice for synthetic data generation. This compound AI system enables generation with complex Jinja templating, Pydantic validation, structured outputs, automated retries, and supports multiple generation backends – the necessary tooling to scale a synthetic dataset of this size. We also leveraged the following models:\n\n1. **Probabilistic Graphical Model (Apache-2.0)** for statistical grounding\n2. **GPT-OSS-120B (Apache-2.0)** for narrative generation in English, Hindi (Devanagari), and Hindi (Latin)\n\n### Embedded Cultural Context\n\nThis dataset was aligned to India’s official demographic distributions from the 2011 Census and expanded to include attributes essential for trustworthy AI training:\n\n- **Education:** Expanded degree levels to reflect India’s diverse academic pathways\n- **Occupations:** Includes formal, informal, and traditional sectors like farming, tailoring, and street vending\n- **Life Stages:** Student, homemaker, retired, and unemployed categories included\n- **Cultural Traits:** Family structures, regional festivals, marriage traditions, and norms\n- **Digital Divide:** Modeled usage patterns across urban/rural, age, and income lines\n- **Linguistic Diversity:** Included incredible diversity with respect to first, second, and third spoken languages for each synthetic persona\n\n### Private By Design\n\nNo real names. No re-identification risk.\n\nAll personas are fully synthetic. While grounded in real-world distributions from the [2011 Census](https://censusindia.gov.in) and [Parsed Indian Electoral Rolls](https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/MUEGDT) data, no data is tied to any living or deceased individual. This ensures developers can safely train AI systems without privacy risks or regulatory barriers.\n\n## Who This Is For\n\n**Built for India, Ready for the World**\n\nNemotron‑Personas‑India is designed for developers building [Sovereign AI](https://www.nvidia.com/en-us/lp/industries/global-public-sector/sovereign-ai-technical-overview/) systems for the **Indian market**, as well as global teams looking to adapt models to India’s unique linguistic, cultural, and social context.\n\nMost open datasets today reflect English-speaking, Western norms—limiting AI performance in India’s multilingual, multi-script, and demographically complex environments.\n\n## Practical AI Applications\n\nWith Nemotron‑Personas‑India, teams can:\n\n- Generate diverse, realistic training data in **Indian languages and scripts**\n- Fine-tune models to capture **local social, occupational, and cultural nuance**\n- Build **region-aware AI agents** that generalize across India’s many communities\n- Develop **domain-specific copilots** tuned to Indian professional and civic workflows\n- Create **multilingual systems** capable of handling **complex multi-turn conversations** and varying levels of digital fluency\n\n## Why It Matters\n\nIndia's 1.4 billion people speak hundreds of languages and live across vast cultural, economic, and geographic divides. India's National AI Portal estimates over [7,000 AI startups and research institutions](https://indiaai.gov.in/indiaaiportal) are working to build locally relevant AI systems, and the Digital India initiative and government programs like IndiaAI are accelerating adoption.\n\nBut progress is constrained by a fundamental gap: high-quality, culturally grounded training data that reflects India's demographic reality. Without representative datasets, AI systems struggle with code-switching between English and Hindi, fail to understand regional occupational categories, and miss cultural context essential for trust and adoption.\n\nThe dataset improves diversity of synthetically-generated data, mitigates biases, and prevents [model collapse](https://medium.com/data-science/addressing-concerns-of-model-collapse-from-synthetic-data-in-ai-7cd380208d14) (degradation caused by uncurated training on another model's outputs) by reflecting India's real geographic and demographic distributions.\n\n**Nemotron-Personas-India** supports Indian model builders in developing Sovereign AI systems that incorporate important region-specific demographics and cultural context.\n\n## Start Building with Nemotron-Personas-India\n\nWant to build AI systems that understand India's culture, languages, and people?\n\n**To start experimenting today:**\n\n```python\nfrom datasets import load_dataset\n\n# English personas\nnemotron_personas_en = load_dataset(\"nvidia/Nemotron-Personas-India\", \"en_IN\")\n# Hindi personas in Devanagari\nnemotron_personas_hi_deva = load_dataset(\"nvidia/Nemotron-Personas-India\", \"hi_Deva_IN\")\n# Hindi personas in Latin\nnemotron_personas_hi_latn = load_dataset(\"nvidia/Nemotron-Personas-India\", \"hi_Latn_IN\")\n```\n\nWhether you're an Indian model builder developing Sovereign AI or a global developer seeking better regional adoption, Nemotron-Personas-India provides the authentic, privacy-safe foundation your applications need.\n\nDownload it. Fine-tune it. Build AI that understands India. If you’re ready to go deeper, an extended version of Nemotron-Personas-India (which includes e.g., first/last names, religion, and synthetic addresses) is available in [NeMo Data Designer](https://docs.nvidia.com/nemo/microservices/latest/generate-synthetic-data/index.html).\n", "dateCreated": "2026-10-04T12:50:30.804479567+00:00", "date_created": "2025-10-13T23:00:42+00:00", "date_modified": "2025-10-13T23:00:42+00:00", "date_published": "2025-10-13T23:00:42+00:00", "description": "A compound AI approach to Indian personas grounded in real-world distributions Open Data for India's AI Future India represents one of the world's largest AI opportunities — with over 700 million internet users, a multitude of languages,…", "external_url": "https://huggingface.co/blog/nvidia/nemotron-personas-india", "git": { "commit": "ebb0941c5c06a2ab2377342d607535262a3e969e", "path": "items/huggingface-co/2025/10/2025-10-13-nemotron-personas-india-synthesized-data-for-sovereign-ai.md", "permalink": "https://github.com/aymericbeaumet/aggr-instance/blob/ebb0941c5c06a2ab2377342d607535262a3e969e/items/huggingface-co/2025/10/2025-10-13-nemotron-personas-india-synthesized-data-for-sovereign-ai.md" }, "id": "urn:aggr:item:4656e1e53d8cad3bd7cb3315654b8aa1d9734c5e", "isBasedOn": { "@id": "https://huggingface.co/blog/nvidia/nemotron-personas-india" }, "mainEntity": { "@id": "https://huggingface.co/blog/nvidia/nemotron-personas-india", "@type": "CreativeWork", "archivedAt": { "@id": "https://aggr.aymericbeaumet.com/items/huggingface-co/2025-10-13-nemotron-personas-india-synthesized-data-for-sovereign-ai/#webpage" }, "dateModified": "2025-10-13T23:00:42+00:00", "datePublished": "2025-10-13T23:00:42+00:00", "headline": "Nemotron-Personas-India: Synthesized Data for Sovereign AI", "timeRequired": "PT5M", "url": "https://huggingface.co/blog/nvidia/nemotron-personas-india", "wordCount": 1020 }, "name": "Archived snapshot: Nemotron-Personas-India: Synthesized Data for Sovereign AI", "preview": null, "reading_minutes": 5, "replicated_at": null, "source": "huggingface-co", "source_name": "Hugging Face - Blog", "summary": null, "tags": [], "temporalCoverage": "2026-10-04T12:50:30.804479567+00:00", "title": "Nemotron-Personas-India: Synthesized Data for Sovereign AI", "url": "https://aggr.aymericbeaumet.com/items/huggingface-co/2025-10-13-nemotron-personas-india-synthesized-data-for-sovereign-ai/", "word_count": 1020 }