satyajitdas/bharatschemes-v1

BharatSchemes Dataset v1

India's first comprehensive instruction-tuning dataset for government welfare schemes, state programs, and corporate CSR/NGO initiatives.

Overview

This dataset contains 2029 instruction-following Q&A pairs covering:

Category	Coverage
Central Government Schemes	MyScheme.gov.in (3000+ schemes)
State Government Schemes	All 28 states + 8 UTs
Corporate CSR Programs	Tata, Infosys, Wipro, Reliance, Azim Premji, Adani, Mahindra
NGO Initiatives	Pratham, Akshaya Patra, GiveIndia, Teach For India, CRY, Goonj
Multilateral Programs	UNDP India

What Makes This Unique

1. Persona-based queries: "I'm a 30-year-old woman farmer in Rajasthan. What schemes can help me?"
2. Cross-state comparisons: "Compare health insurance schemes across states"
3. Application guidance: Step-by-step process for each scheme
4. Unified coverage: Central + State + CSR + NGO in one dataset (first of its kind)

Format

ChatML format (compatible with MLX, Axolotl, LLaMA-Factory):

{
  "messages": [
    {"role": "system", "content": "You are BharatSchemes Assistant..."},
    {"role": "user", "content": "What is PM-KISAN?"},
    {"role": "assistant", "content": "PM-KISAN is..."}
  ]
}

Splits

• train.jsonl: 1927 examples
• valid.jsonl: 102 examples

Usage

from datasets import load_dataset
ds = load_dataset("satyajitdas/bharatschemes-v1")

Author

Satyajit Das (IIT Kharagpur alumnus)

• Cambridge PhD in Engineering
• Consensus Engineer @ Offchain Labs (Ethereum / Prysm)
• Protocol Guild Member
• Publication: Science (nano architecture, ~1000 citations)

Citation

@dataset{bharatschemes2026,
  title={BharatSchemes: Instruction Dataset for Indian Welfare Scheme Navigation},
  author={Das, Satyajit},
  year={2026},
  url={https://huggingface.co/datasets/satyajitdas/bharatschemes-v1}
}

License

Apache 2.0. Source data from public government websites.