SARTHI AgriData

The Agri-Advisory Synthetic Dataset (Hindi) is a large-scale instruction-tuning dataset for Indian agriculture, containing 2,20,222 biologically validated advisory examples. Generated through a multi-stage pipeline separating scientific feasibility from language generation, each record includes a structured agricultural scenario, a Chain-of-Thought reasoning trace, and a final Hindi advisory. The dataset supports chatbots, IVR, and radio-style delivery.

About Dataset

The Agri-Advisory Synthetic Dataset (Hindi) is a large-scale, high-quality dataset with 2,20,222 agricultural advisory examples designed for training LLMs in the Indian farming context. Each record includes a scientifically validated crop scenario, the model’s reasoning (Chain-of-Thought), and a clear, actionable Hindi advisory. The dataset is generated using a multi-stage pipeline that filters out biologically invalid scenarios before text generation, ensuring reliability and safety. It supports multiple advisory styles (chat, radio, IVR) through persona-based instructions and enforces strict guardrails for organic and conventional practices. Available in JSONL and Parquet, it is ideal for instruction tuning, reasoning-focused training, and real-world agricultural AI applications.

Purpose of Dataset

Utility, Use Cases, and Policy Relevance The Agri-Advisory Synthetic Dataset (Hindi) has strong practical utility across research, deployment, and governance of agricultural AI systems in India. Utility & Use Cases LLM Training for Agri-Advisory Systems: Enables instruction tuning and reasoning-aware training of models that provide crop-, region-, and stage-specific advice in Hindi, improving accuracy and farmer trust. Farmer-Facing Applications: Suitable for chatbots, WhatsApp advisories, IVR systems, and community radio broadcasts due to persona-driven outputs and TTS-friendly formats. Decision Support for Extension Services: Can assist Krishi Vigyan Kendras (KVKs), agri-extension officers, and NGOs with consistent, localized advisory content. Reasoning & Safety Research: The inclusion of Chain-of-Thought allows evaluation of model reasoning quality, robustness, and hallucination reduction in high-stakes domains. Low-Resource Language Enablement: Strengthens Hindi agricultural NLP resources, addressing a major gap in non-English, domain-specific datasets. Policy Relevance Alignment with Digital Agriculture Initiatives: Supports national programs such as Digital Agriculture Mission, AgriStack, and AI-based farmer advisory platforms. Safe & Responsible AI: Pre-validation of biological feasibility and enforced guardrails (organic vs. chemical, PPE warnings) align with emerging AI safety and responsible deployment guidelines. Scalable Public Advisory Infrastructure: Enables cost-effective, scalable dissemination of scientifically grounded advisories to small and marginal farmers. Evidence-Based Policymaking: Can be used to simulate and test advisory policies, stress scenarios, and climate-impact responses before real-world rollout. Overall, the dataset bridges scientific agricultural knowledge with trustworthy AI generation, making it valuable for both operational agri-tech systems and policy-driven digital agriculture initiatives in India.