abhirud/AgriAssist-2B

Farmer-Query-Assistant-India

Farmer Query Assistant India is a lightweight instruction-tuned language model developed to provide agricultural guidance for Indian farmers. Built upon BeejX-Gemma2-2B-Hindi-SFT, the model has been further fine-tuned using QLoRA with the Unsloth framework on real-world agricultural conversations from India. The model is distributed in GGUF format for efficient local inference on resource-constrained systems using modern GGUF-compatible runtimes.

Model Details

• Base Model: BeejX-Gemma2-2B-Hindi-SFT
• Fine-tuning: Supervised Fine-Tuning (SFT) using QLoRA on the DigiGreen/farmerchat-queries-large dataset, filtered to include only queries originating from India.
• Languages: Hindi and English.
• Training Framework: Unsloth + Hugging Face Transformers + TRL.
• Format: GGUF (Q4_K_M quantization).

Capabilities

• Provides recommendations for crop cultivation and farm management.
• Assists with pest, disease, and nutrient deficiency related queries.
• Offers guidance on irrigation, fertilizer application, and general agricultural practices.
• Utilizes regional information provided in the prompt to generate more context-aware responses.
• Generates concise, instruction-following responses suitable for conversational agricultural assistance.

Intended Applications

• Agricultural advisory assistants
• Offline AI assistants
• Local deployment with Ollama or llama.cpp
• Educational and research projects in agricultural AI
• Integration into farming support applications and chatbots

Running the Model

This model is distributed in GGUF format and can be used with any GGUF-compatible inference engine.

Supported runtimes include:

• Ollama
• llama.cpp
• LM Studio
• Jan
• Open WebUI

1. Download the desired .gguf file from the Files section of this repository.
2. Load the model using your preferred GGUF runtime.
3. A context size of 2048 or higher is recommended for best results.

Recommended System Prompt

You are an AI agricultural assistant specializing in Indian farming practices.

Provide practical, accurate, and easy-to-understand guidance for agricultural questions. Consider the user's crop, symptoms, and location whenever available. Answer in the same language as the user's query whenever possible. If sufficient information is not provided, ask relevant follow-up questions before making recommendations. Avoid fabricating facts or providing unsafe agricultural advice.

Limitations

• The model is intended to provide informational guidance and should not replace recommendations from certified agricultural experts or local agricultural extension services.
• It does not have access to real-time weather, market prices, or government databases.
• Responses are limited by the coverage and quality of the supervised fine-tuning dataset.