Named after Kamadhenu, a divine bovine-goddess described in Hinduism as the mother of all cows, the model is trained on high-quality conversational datasets, specially focused on Indian agriculture practices. Dhenu is an agricultural LLM specifically designed to help Indian farmers with agriculture-related questions.
Designed to be bilingual, it has been trained on 300,000 instruction sets in English and Hindi.
“Our goal was to develop an AI/ML application that can be used by farmers. Throughout our journey we have experimented with building different tools and have been expanding our knowledge base,” Desai told ET in an interaction on January 3.
In March last year, the company launched KissanAI, a voice-to-voice end-to-end chatbot for farmers in 10 Indian languages. Farmers could talk in their own language and get queries answered on agriculture inputs, package of practices for different crops, fertilisers and pesticides.
Desai realised that for his solution to be scalable and reach farmers, it had to be cheap even though the cost of running on GPUs was expensive for him.
Discover the stories of your interest
“Our goal was to build a small model based on more than 355,000 agriculture conversations we had had with more than 100,000 farmers, and voice or text datasets we had collected on our KissanAI platform,” he said.Sarvam AI
After AI startup Sarvam AI unveiled its Hindi LLM OpenHathi, KissanAI collaborated with them. The datasets were trained on OpenHathi and fine-tuned. “This reduced our cost four times. It also helped in terms of latency,” he said. Desai plans to add other language datasets to his model gradually.
Agriculture datasets
Some of the datasets were sourced from pictures of booklets or pamphlets with agricultural information. So, the firm used optical character recognition to digitise the text and worked on data annotation and curation. They collaborated with three agricultural universities to get the terminologies correct.
KissanAI also signed an agreement with Navsari University in Gujarat for knowledge transfer. They also sought help from those working in a university in Jamnagar and Anand Agricultural University for doubts in areas like animal husbandry or horticulture. They collected open data from 30 academic institutes.
ETtechSeveral acronyms had to be expanded to proper names of seeds, fertilisers and pesticides for the data to be used for training. “Geographical diversity was another challenge. A question about paddy from Kerala may have different answers for other parts of India due to the difference in practices,” he explained.
Currently, Dhenu is under testing and evaluation. “When we go from 300,000 to a million instructions, we are confident of having a high-quality model and agriculture companies can integrate with this model,” he said.
The problem statement
Every agricultural university has a mandate to provide a package of practices for different varieties of crops in their area to be given to the farmers. Hence, they create booklets or printouts with the information. The biggest problem is illiteracy among farmers. They cannot read or they do not have access to universities to read these booklets and solve their problems.
“We collect them and proofread them. We process, contextualise and correct the semantics, and translate them. This can be a standard advisory on seed preservation,” he said.
Usage
Desai said that he will create a premium model of Dhenu and provide limited access. “We will open source the current version and train a premium version and create an API-usage based model,” he said. So, if one wants to integrate with the model, they may use the premium version.
Once Dhenu is ready, it will replace the GPT 3.5 that is currently the base of KissanAI chatbot.
“We will add market analysis to the KissanAI interface so that farmers can start directly asking via voice queries. What is the average price in the nearest mandi? What was the price last month? We can even add weather data to this. They can enquire about policies, schemes and equipment details. This is how the chatbot experience will be expanded,” he explained.
Although Dhenu is about India-specific agriculture, Desai is already getting queries from Africa, South America and Indonesia.
“Our goal is to start from here and start having versions for different countries. If this can work for India’s geography, diversity, weather, climate and languages, then mostly it will work for other countries that are monolithic. Agriculture datasets in other countries like the US may be more digitised compared to our country,” he said.
Partnership with UNDP
KissanAI, on January 5, announced partnering with the United Nations Development Programme to develop a vernacular co-pilot for climate resilient agriculture practices targeting rural, smallholder and women farmers.
“We are excited to bring generative AI to remote corners of India and then the world. We are targeting rural and mostly illiterate farmers so voice in their language can enable them to get information. The goal is to focus on women farmers,” Desai told ET.
Background
Son of a farmer, Desai heads a technology team with engineers, those who do data curation and data science, and almost all members have an agriculture background.
“One common thread that binds us is an agriculture background. Barring one, all our fathers are farmers. The benefit is to understand any part of agriculture, we can just call our fathers,” Desai said.
He said until he turned 21 and moved to the US to pursue his post graduate programme and PhD, he helped his father farm. “We used to grow tur dal, bajra, sugarcane and castor in Surat. Our farms are in Surat and so is our technology team,” he said.
He worked on three startups before jumping into agriculture technology. Since 2019, he has been working full time on Indian agriculture technology while being based in San Francisco along with his family.