AgriParam: a bilingual language model for Indian agriculture

Agriculture is India's largest employer, yet most language models are built for global, English-speaking audiences and generic use cases. They know little about Indian crop varieties, soil types, agro-climatic zones, seasonal calendars or government schemes, and they rarely work well in Hindi.

AgriParam is our answer to this gap. It is a 2.9 billion parameter language model fine-tuned from Param-1-2.9B-Instruct on an India-centric agriculture dataset. AgriParam answers agricultural queries in English and Hindi, and is designed for farmer advisories, policy and scheme information, research insights and rural knowledge sharing.

Under the hood: AgriParam's architecture

AgriParam inherits the decoder-only transformer architecture of Param-1-2.9B-Instruct and adapts it to agriculture through supervised fine-tuning.

Technical specifications:

• Parameters: 2.9 billion
• Architecture: Decoder-only transformer with 32 layers, hidden size 2,048 and intermediate size 7,168
• Attention: Grouped-query attention with 16 heads and 8 key-value heads
• Context length: 2,048 tokens
• Embeddings: Rotary positional embeddings (RoPE) with theta=10,000
• Activation: SiLU
• Precision: bf16-mixed
• Vocabulary: 256,000 tokens plus four special tokens (<user>, <assistant>, <context> and <system_prompt>) for clean instruction formatting

Data preparation: grounding the model in Indian agriculture

A useful agriculture model needs the right data first. AgriParam's training corpus was built in four steps:

1. Source gathering: About 17,000 open-source, India-focused agricultural news articles and knowledge passages were scraped and curated.
2. Question generation: An open-weight LLM generated five question-answer pairs for each passage.
3. Taxonomy and personas: We built an India-specific agricultural taxonomy spanning crops, weather, soil, pests, fertilisers, market conditions and government schemes. We also defined personas such as smallholder farmers, agri-extension officers, policymakers, scientists and agri-business users.
4. Dataset construction:
   • 2 million question-answer pairs grounded in the taxonomy and personas
   • The complete dataset translated into Hindi for bilingual coverage
   • 6 million multi-turn conversations simulating agricultural advisory dialogues in rural settings

The result is a bilingual, domain-aligned dataset that captures not just agricultural facts but also the intent and context of how farmers and advisers ask questions in India.

Training methodology

AgriParam was trained with supervised fine-tuning, using a custom prompt template designed for multi-turn, context-rich agricultural queries.

Training configuration:

• Base model: Param-1-2.9B-Instruct
• Framework: Hugging Face with multi-node torchrun distributed training
• Training samples: 12 million
• Test set: 1.2 million examples
• Training epochs: 3

Optimisation parameters:

• Learning rate schedule: Linear with warmup
• Base learning rate: 5e-6
• Batch configuration: Global batch size of 1,024 with a micro batch size of 4
• Gradient accumulation: 32 steps

The model supports three prompt formats: general question answering, question answering over a supplied context, and multi-turn conversations of up to five turns.

How AgriParam is evaluated: BhashaBench‑Krishi

General benchmarks do not test whether a model can give useful, localised agricultural advice. AgriParam is evaluated on BhashaBench-Krishi (BBK), our benchmark built from real agricultural examinations in India. Unlike generic tests of factual recall, BBK checks whether a model has the practical, region-aware knowledge Indian farmers and advisers rely on. You can read more in our BhashaBench-Krishi write-up.

Assessment scope:

• Question volume: 15,405 validated questions from 55+ government and institutional agricultural exams, from NABARD Grade A and MP RAEO to ICAR postgraduate and PhD entrance exams
• Domain coverage: 25+ agricultural and allied domains spanning more than 270 topics, from soil science to government schemes
• Language distribution: 12,648 English and 2,757 Hindi questions
• Difficulty stratification: Easy (6,754), Medium (6,941) and Hard (1,710) questions
• Question formats: Multiple choice, assertion-reasoning, match the column, rearrange the sequence and fill in the blanks

What BBK tests that generic benchmarks miss:

• Policy navigation: Government schemes and subsidy eligibility
• Regional adaptation: State-specific practices and crop calendars
• Seasonal knowledge: Advice suited to the kharif, rabi and zaid seasons
• Practical application: Turning scientific knowledge into farming decisions

Performance analysis

AgriParam records the highest overall BBK score among the similar-sized models we compared, although the margins at the top are small.

AgriParam scores 32.18% overall (33.10% in English and 27.97% in Hindi), narrowly ahead of Granite-3.1-3B-A800M (32.17%) and Llama-3.2-3B (31.96%). In Hindi alone, Llama-3.2-3B scores slightly higher (28.69%).

Domain strengths: AgriParam scores highest among the compared models in these domains:

• Veterinary Sciences: 43.75%
• Animal Sciences: 40.54%
• Soil Science: 34.93%
• Agricultural Extension Education: 34.50%
• Agronomy: 32.49%, in the largest domain in BBK with 5,078 questions

Performance by question difficulty:

• Easy questions: 36.94%, the highest among the compared models
• Medium questions: 29.09%
• Hard questions: 25.91%

Where AgriParam needs work: The model is weaker in some specialised domains, including Fisheries and Aquaculture (23.53%), Nematology (23.91%) and Plant Pathology (25.44%), and on hard questions. These are priorities for future versions.

Why this matters

Most AI tools available to farmers today are general-purpose models with agricultural terms layered on top. AgriParam takes a different route: a compact, sector-specialised model trained on Indian agricultural knowledge in English and Hindi.

It can serve as a base for Kisan helplines, chatbots for Krishi Vigyan Kendras, voice assistants for rural users and advisory tools for extension workers. AgriParam is our first step towards domain-native language models that serve real people with real needs.

Resources:

• Model access: AgriParam on Hugging Face
• Benchmark dataset: BhashaBench-Krishi
• Base model: Param-1-2.9B-Instruct
• Technical documentation: Available in the model repository

Want to partner with us, contribute, or share feedback on AgriParam? Get in touch with our team.