Sarvam AI, Indian Startup, Launches First Open-Source Hindi AI Model
Sarvam AI, an Indian startup, has launched OpenHathi-Hi-0.1, marking the introduction of an open-source Hindi language model. This launch initiates a series aimed at encouraging innovation in Indian language AI by contributing open models and datasets to the ecosystem. Built upon Meta AI’s Llama 2-7B model, Sarvam AI's blog claims that this model is on par with GPT-3.5 for Indic languages. Overcoming the challenge of tokenization, particularly expensive in Hindi due to limited training text, was addressed through cost-effective methods during the model’s two-phase training.
Testing involved a range of assessments, spanning from traditional benchmarks like translation to innovative evaluations such as toxicity checks and text classification. The foundational model is now accessible on the Hugging Face platform, allowing developers to fine-tune and deploy it for specific purposes. Pratyush Kumar and Vivek Raghavan, Co-Founders previously associated with AI4Bharat, collaborated with the organization, utilizing language resources and benchmarks to train OpenHathi.
With around 18 team members, Sarvam AI aims to create comprehensive language models integrating voice as a universal interface, tailored to meet the diverse needs of the Indian market. Having secured $41 million in Series A funding, primarily led by Lightspeed Ventures with contributions from Peak XV and Khosla Ventures, the startup, only five months old, continues to make significant progress.
