Cropin’s AI model to aid farmers
Google-backed agritech startup Cropin Technology unveiled its latest initiative on Tuesday: an open-source micro language platform named ‘Aksara’, built on Mistral’s foundation model.
The solution is designed to enhance major crop production in South Asia, with a focus on sustainable and energy-efficient farming practices while offering comprehensive language support.
Krishna Kumar, co-founder and chief executive of Cropin, said the goal is to democratize access to digital technologies as well as modernize agricultural processes.
Cropin aims to empower agricultural researchers and developers to tackle global challenges like food security, climate change, resource conservation (water and soil), and regenerative agriculture practices by offering access to contextual, factual, and actionable information.
Initially, it will cover nine crops, including paddy, wheat, maize, sorghum, barley, cotton, sugarcane, soybean, and millet, spanning five countries in Indian subcontinent. These food crops account for a substantial portion of the world’s food requirements and are essential staples for the population in the global south, Kumar said.
Praveen Pankajakshan, vice president of data science and AI, Cropin, said the technology is both cost-effective and scalable, built and fine-tuned on top of the Mistral-7B-v0.2 model, developed by Cropin and hosted on Hugging Face.
It compressed ‘Aksara’ from 16-bit to 4-bit, utilising quantization and low-rank adapters to reduce the environmental impact of running large language models. The model outperforms GPT-4 Turbo by 40% on randomly selected test datasets, as measured by the ROUGE scoring algorithm, said Pankajakshan. It ensures that the responses are factually relevant and brief while minimizing computing and storage resource requirements.
The model was fine-tuned with over 5,000 high-quality question-response pairs specific to agriculture and over 160,000 in-context tokens. These numbers are expected to increase as more geographic locations, crops and use cases are added, said Pankajakshan. The model is faithful to questions by using techniques like retrieval augmented generation through cross-referencing expert knowledge bases.