Google I/O 2024 concluded recently with an array of noteworthy announcements. In the latter part of the keynote, Google presented a brief video highlighting Gemma’s capabilities in constructing Indic LLMs.
The tech giant unveiled Navarasa, a Gemma 7B/2B instruction-tuned model accommodating 15 Indian languages alongside English. This innovation was spearheaded by Telugu LLM Labs, established by Ravi Theja Desetty and Ramsri Goutham Golla.
Navarasa’s training took place on E2E Networks Limited, utilizing NVIDIA A100 GPUs. The process consumed around 44 hours for the 7 billion model and 18 hours for the 2 billion model.
Harsh Dhand, head of APAC research partnerships at Google, emphasized the significance of tailoring technology to specific cultures like India for nuanced understanding and problem-solving.
Take your side projects seriously!
You might as well get featured in the Google I/O keynote on the main stage in California 😉
I’m excited to share that the work @ravithejads and I did in building Navarasa, an Indic instruction-finetuned model on top of Google’s Gemma catering… pic.twitter.com/8B1I0p3iYl
— Ramsri Goutham Golla (@ramsri_goutham) May 15, 2024
AIM approached Theja and Golla for insights into their involvement. Golla shared their experience, noting Google’s generous arrangements during their filming stint in Mysore, which made them feel like celebrities.
Theja revealed that Google discovered their work through his blog post about Navarasa, expressing surprise at its showcase during the main event of Google I/O.
Golla, who spent almost eight years studying and working in the US before returning to India, described himself as a builder/engineer with a penchant for creating SaaS apps. Meanwhile, Theja, a developer advocate engineer at Llama Index, previously served as a senior ML engineer at Glance, focusing on recommendation systems and GenAI applications.
Navarasa 2.0, a Gemma 7B/2B SFT model, expands generative capabilities across 15 Indian languages, achieved by translating datasets into six additional languages.
The model is accessible on IndicChat, a platform facilitating interaction with Indic AI models. Golla highlighted Gemma’s advantage over Llama in terms of fine-tuning requirements and token support for various languages.
OpenAI’s GPT-4o has significantly enhanced support for Indian languages, reducing token usage and extending vocabulary size, as noted by Abhishek Upperwal of Socket AI Labs.
Users have praised OpenAI’s attention to Indic languages, although Theja expressed concerns about the audio feature’s performance in Indian languages.
Indian AI startup Sarvam AI is developing an Indic Voice LLM, while Hanooman and Krutrim are expected to incorporate voice capabilities in the near future.
