KissanAI’s founder, Pratik Desai, has had a hectic few weeks, working against a tight timeline to unveil Dhenu 1.0, India’s first agricultural LLM. That too before the end of December, as he had to present the model at the GPAI Summit 2023 which took place in India from 12-14 December 2023.
“For seven days, I had 24-hour flights, and for the remaining three days, I was stranded at the airport the whole time.” he said, sharing his experience of showcasing the first version of Dhenu at GPAI Summit, in an exclusive interview with AIM.
Dhenu is an agricultural LLM specifically designed for Indian farmers to assist them with queries related to farming practices. The model is bilingual and can comprehend queries in English, Hindi, and Hinglish, a notable feature that caters directly to farmers’ linguistic needs.
Collaborates with Sarvam AI
The first version of Dhenu is built on top of Sarvam AI’s recently launched model, OpenHathi. Desai said that he tried to fine-tune Dhenu earlier on Mistral 7B, but due to the lack of time (as he had to present his model at the GPAI Summit), he collaborated with Sarvam AI.
The partnership with Sarvam AI is primarily about compute. “We were in coordination with Sarvam AI, and as soon as OpenHathi became available, we decided to use it,” said Desai. He mentioned that he knew folks at the company beforehand, so collaborating with them was easy.
Moreover, the reason for choosing OpenHathi was that it is a bilingual model, addressing KissanAI’s major problem. “The intention was also to create a Hindi LLM so that Indian farmers could easily use it.” he said.
The Art of Data Collection
Desai said, unlike general-purpose models built to answer generic queries and create witty responses, Dhenu 1.0 serves a different purpose. He mentioned that his team has collected agricultural data from various universities across India, spanning from the north to the south.
Furthermore, he highlighted that Kissan AI has recorded real conversations related to agriculture, providing genuine answers. “We have enough information to start training, and that is making it very domain-specific now,” he added.
The model’s uniqueness lies in its bilingual nature, adeptly processing 300,000 instruction sets in both English and Hindi according to Desai. “300,000 instructions contain Indian agricultural data practices, backed by Indian methods. He mentioned that the model will continue to train on more data in the coming months.
“We have also incorporated geographical information into this instruction set, indicating the origins of the questions,” said Desai. Furthermore, he said that KissanAI has a significant amount of Indic knowledge data, such as literal pictures taken by some Krishi Vigyan Kendra individuals or written documentation by professors.
Subsequently, these materials become instructions for farmers, presented in their own language. “There is a wealth of knowledge traditionally stored in pen-and-paper format, drawings, or manuals that we aim to bring online, convert into AI, and make accessible.” added Desai.
KissanAI’s Business Model
Still in the early days, Desai was quite vocal, and said he hasn’t figured out the business model yet. “I don’t know the strategy right now. I am still thinking about it. There can be an open-source version, along with a pro version,’ he said, explaining that he definitely has plans to sell APIs directly to agriculture companies aiming to build their own co-pilots for agriculture.”
Moreover, Dhenu is not yet available in the market as it is currently undergoing evaluation. “The goal is to have some RAG testing with the model. And, evaluation may take time because we are a small team,” shared Desai.
Desai told AIM that he has plans to take Dhenu outside India as well. He said that some of the companies in Brazil reached out to him to create a similar model for them.
He mentioned that it is easier to build models for other nations as they are more monolithic, unlike India, which has such diversity in terms of language and culture, alongside agricultural practices, as the majority are marginal farmers.
Lastly, when asked about the choice of the name ‘Dhenu,’ he shared “I think it started with people naming all the models with animal names like LLaMA or Hyena. So, we wanted to have a name that follows a sequence because, ultimately, we are a community building on each other’s work.” He said that naming their company Kissan and their model Dhenu makes sense, as both are closely related to agriculture.