It was built on Mistral’s seven billion parameter foundation model and fine-tuned on data from Cropin’s internal knowledge base, specific to nine major crops – paddy, wheat, maize, sorghum, barley, cotton, sugarcane, soybean, and millets – on conditions in India, Pakistan, Bangladesh, Sri Lanka and Nepal. It used over 5,000 question-response pairs specific to agriculture and over 160,000 tokens in the context.
Cropin has raised a total of $47 million in funding. Of the $34 million raised in the last two rounds, around 40% of capital has gone towards advancing intelligence and building assets for scale, Sujit Janardanan, chief customer and marketing officer, Cropin, told ET.
The Bengaluru-headquartered company counts Google, Chiratae Ventures, the Bill & Melinda Gates Foundation, and Infosys cofounder Kris Gopalakrishnan’s family office Pratithi Investment Trust among its investors.
Founded in 2010, its intelligent agriculture solutions are deployed in over 103 countries.
The company’s business has been doubling year-on-year, founder and CEO Krishna Kumar said, without disclosing specific revenue numbers. He added that a micro model was chosen as it is more cost-effective and energy-efficient compared to large language models. Compressed from 16-bit to 4-bit, the model employs minimal compute resources.
Discover the stories of your interest
“The thinking behind it is how do we show the path to a greener world, where while we want to solve the problem with AI, we don’t create another problem of energy or carbon footprint,” Kumar said.The model can suggest what should be done at the time of seeding, appropriate nursery preparation techniques, what practices would help to conserve water etc.
While the first version is limited to English, in the coming six months, the model will be fine-tuned for a wider range of languages, geographies, and crops.
“Right now, it’s more for researchers, enterprises, developers to download and play around with, release some models, embed within their applications, take it to their agronomies and to the fields. Slowly, it can go in the hand of growers as well, as and when we bring the voice capability and then language capability in this application,” Kumar said.
The company said that Aksara was found to be more relevant than GPT-4 Turbo by almost 40% on randomly selected test datasets, measured by the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) scoring algorithm.