AI4Bharat, an Indian AI research lab incubated at IIT Madras, has introduced Airavata, a novel instruction-tuned model dedicated to the Hindi language.
Airavata’s inception is the result of AI4Bharat’s collaboration with a research team led by Mitesh M. Khapra, Assistant Professor at IIT Madras and Co-Founder of AI4Bharat. The team meticulously crafted the model using Sarvam AI’s OpenHathi, a comprehensive dataset, to fine-tune it for instruction-based tasks. This approach is crucial for effective language understanding and generation, making Airavata a significant tool in various applications.
The uniqueness of Airavata lies in its instruction tuning, a method that tailors the model to understand and follow specific instructions in Hindi. Unlike other models that rely on broad language datasets, Airavata’s instruction-tuned approach ensures higher accuracy and relevance in processing Hindi text.
One of the critical challenges in the development of Airavata was the lack of extensive Hindi datasets for instruction tuning. To overcome this, AI4Bharat employed a novel strategy of translating existing English instruction-tuning datasets into Hindi. This was accomplished using IndicTrans2, a cutting-edge machine translation model designed for Indian languages.
AI4Bharat’s vision extends beyond Airavata. The lab has previously launched impactful projects like Chitralekha, an open-source AI-powered video transcreation platform, developed in collaboration with EkStep. Furthermore, the lab has announced its AI resident and associate program for 2024-25, aiming to cultivate expertise in fields like NLP, speech, and vision.
The release of Airavata by AI4Bharat is a significant stride in bridging the language gap in AI technology as the current trend has been skewed towards English-centric models. By focusing on Hindi, a language spoken by a vast population yet underrepresented in the AI domain, Airavata opens new avenues for technology to serve and understand a broader demographic.