BharatGPT group unveils ‘Hanooman’: Everything you need to know about the Indic AI model

What is Hanooman?

Essentially, Hanooman is a series of large language models (LLMs) that can respond in 11 Indian languages like Hindi, Tamil, and Marathi, with plans to expand to more than 20 languages. According to a Bloomberg report, BharatGPT group in a video on Tuesday, showed different people interacting with the AI tool in different languages.

Hanooman has been designed to work in four fields, including health care, governance, financial services, and education.

Notably, the series isn’t just a chatbot. It is a multimodal AI tool, which can generate text, speech, videos and more in multiple Indian languages, according to BharatGPT. One of the first customised versions is VizzhyGPT, an AI model fine-tuned for healthcare using reams of medical data.

The size of these AI models ranges from 1.5 billion to a whopping 40 billion parameters.

Vishnu Vardhan, the Founder of SML, during the launch of Hanooman, noted the challenges posed by the quality of datasets in Indian languages. He highlighted the prevalence of synthetic datasets — information that’s artificially generated instead of produced by real-world events — derived from translations, which could lead to inaccuracies or distortions, a report by ANI news agency said.

Are there any other Indian language models?

Apart from BharatGPT, a host of different startups like Sarvam and Krutrim, backed by prominent VC investors such as Lightspeed Venture Partners and billionaire Vinod Khosla’s fund, are also building AI models customised for India, according to the Bloomberg report.

What are LLMs?

Large language models use deep learning techniques to process large amounts of text. They work by processing vast amounts of text, understanding the structure and meaning, and learning from it. LLMs are ‘trained’ to identify meanings and relationships between words. The greater the amount of training data a model is fed, the smarter it gets at understanding and producing text.

The training data is usually large datasets, such as Wikipedia, OpenWebText, and the Common Crawl Corpus. These contain large amounts of text data, which the models use to understand and generate natural language.