Nemotron-Personas-India: Synthesized Data for Sovereign AI

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, 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.

Today, 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.

Built 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.

This 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.

This 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.

What’s in the Dataset?

• 21 million personas total (3M records × 7 personas each)
• Multilingual support: English and Hindi, in both Devanagari and Latin scripts
• 27 fields per record: Persona traits + contextual attributes grounded in official census and labor statistics, including age, gender, education, occupation, state, district, and more
• 7.7 billion tokens total, including 2.9B persona tokens
  • English: 1B tokens total, 394M persona tokens
  • Hindi (Devanagari): 4.7B tokens total, 1.8B persona tokens
  • Hindi (Latin): 2B tokens total, 746M persona tokens
• ~560k unique full names, reflecting India's vast linguistic diversity
• 2.9k occupational categories, including informal, formal, and traditional sectors
• All 36 states of India and 640 districts represented
• Natural language fields cultural background, linguistic background, skills and expertise, hobbies and interests
• Persona types: Includes general, professional, linguistic, culinary, sports, arts, and travel personas
• Licensed under CC BY 4.0 for commercial and non-commercial use

How We Built It

Data Generation Pipeline

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:

1. Probabilistic Graphical Model (Apache-2.0) for statistical grounding
2. GPT-OSS-120B (Apache-2.0) for narrative generation in English, Hindi (Devanagari), and Hindi (Latin)

Embedded Cultural Context

This dataset was aligned to India’s official demographic distributions from the 2011 Census and expanded to include attributes essential for trustworthy AI training:

• Education: Expanded degree levels to reflect India’s diverse academic pathways
• Occupations: Includes formal, informal, and traditional sectors like farming, tailoring, and street vending
• Life Stages: Student, homemaker, retired, and unemployed categories included
• Cultural Traits: Family structures, regional festivals, marriage traditions, and norms
• Digital Divide: Modeled usage patterns across urban/rural, age, and income lines
• Linguistic Diversity: Included incredible diversity with respect to first, second, and third spoken languages for each synthetic persona

Private By Design

No real names. No re-identification risk.

All 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.

Who This Is For

Built for India, Ready for the World

Nemotron‑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.

Most open datasets today reflect English-speaking, Western norms—limiting AI performance in India’s multilingual, multi-script, and demographically complex environments.

Practical AI Applications

With Nemotron‑Personas‑India, teams can:

• Generate diverse, realistic training data in Indian languages and scripts
• Fine-tune models to capture local social, occupational, and cultural nuance
• Build region-aware AI agents that generalize across India’s many communities
• Develop domain-specific copilots tuned to Indian professional and civic workflows
• Create multilingual systems capable of handling complex multi-turn conversations and varying levels of digital fluency

Why It Matters

India'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.

But 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.

The 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.

Nemotron-Personas-India supports Indian model builders in developing Sovereign AI systems that incorporate important region-specific demographics and cultural context.

Start Building with Nemotron-Personas-India

Want to build AI systems that understand India's culture, languages, and people?

To start experimenting today:

from datasets import load_dataset

# English personas
nemotron_personas_en = load_dataset("nvidia/Nemotron-Personas-India", "en_IN")
# Hindi personas in Devanagari
nemotron_personas_hi_deva = load_dataset("nvidia/Nemotron-Personas-India", "hi_Deva_IN")
# Hindi personas in Latin
nemotron_personas_hi_latn = load_dataset("nvidia/Nemotron-Personas-India", "hi_Latn_IN")

Whether 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.

Download 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.