This article is under draft review and may be edited further in the near future.
Himachal Pradesh is an emerging state in natural farming, which focuses on improving soil health, reducing chemical use, and protecting biodiversity. Over 2.23 lakh farmers across all panchayats have already adopted this approach, and the state aims to bring 9.61 lakh farmers on board. This shift promotes environmentally friendly and cost-effective agriculture.
However, farmers are facing a primary challenge of pest management. Currently, we receive pest management advice only after the visible damage has occurred. To address this issue, we need a predictive model that can forecast potential damage before it occurs.
Reducing Hallucinations with Domain-Specific SLMs
The Large Language Models (LLMs) are a powerful way for the analysis and organization of information as well as the ability to interact with technology more naturally.
Models such as ChatGPT have been trained on general internet data, which may mix conventional and natural farming information. This may lead to the generation of wrong information or misleading answers. The major risk is that the models may hallucinate by providing inaccurate information or incorrect outputs confidently. Even fully fine-tuning an LLM will need high computational power, which is costly.
SLM, on the other hand, being lightweight, is easy to retrain or fully fine-tune. This approach allows for a reduction in hallucinations if SLM is trained on domain-specific data.
POC: Fine-Tuning SLM with Natural Farming Data
A proof of concept (POC) is being developed using the small language model, such as the google/gemma-3-1b-it version. The SLM is being trained on domain-specific data related to natural farming. Being trained on this data, the model will hallucinate less and provide more accurate results as compared to any other LLM. The model won’t suggest anything outside of the natural farming domain. Moreover, the model’s context awareness will be better.
The approach currently taken is to convert unstructured data into JSONL format for supervised full fine-tuning of the model. The GitHub code provided is to build the baseline model. Currently the model is being fine-tuned on a limited amount of data, due to which it will hallucinate more, but with the increase in the volume of data in the future, the hallucination will decrease to an extent.
https://github.com/prikshitkverma/Gemma_fine_tuning
From Data to Decisions: SLM–IoT Integration
The crucial step will involve integrating the SLM with the IoT devices that continuously monitor all relevant data and transform the statistical knowledge into a predictive decision support tool. This will help deliver the early warnings of the pest outbreaks.
This will help farmers shift from a reactive approach, where farmers intervene when the damage is visible, to a proactive approach, where early warnings can help to take preventive measures and help reduce crop losses.