MBZUAI open-sources NANDA LLM
10 billion parameter version of Hindi language NANDA LLM released
#India #UAE #LLMs - Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), the world's first graduate-level AI research university, has released the code for Llama-3-Nanda-10B-Chat under Meta’s Llama 3 licence. Developed in collaboration with G42’s applied research arm Inception and Californian AI infrastructure company Cerebras Systems, the Hindi large language model (LLM) NANDA was first announced at the UAE-India Business Forum in Mumbai, India during September. The AI model was developed to enable over half a billion Hindi speakers to use generative AI in their native language.
SO WHAT? - In developing state-of-the-art Arabic language large language models, G42’s Inception and Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) have gained considerable expertise in, and created methodologies for developing LLMs capable in languages that use non-Latin alphabets. Therefore, the departure from developing Arabic LLMs to build a Hindi language model is not a complete surprise. Leveraging this non-Latin language capability, the building of a Hindi LLM was added to a bilateral agreement on digital infrastructure between India's Ministry of Electronics and Information Technology and the UAE's Ministry of Investment earlier this year. The resulting Hindi LLM, NANDA, was launched in India during September.
Here are some key details regarding the LLM:
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has open-sourced the code for Llama-3-Nanda-10B-Chat under Meta’s Llama 3 licence.
The NANDA LLM released is a 10 billion parameter version optimised for chat, following the launch of the NANDA 13 billion parameter Hindi-centric large language model in September.
NANDA was developed by G42 group’s applied research arm Inception, in collaboration with the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Cerebras Systems.
The 13 billion parameter NANDA LLM was trained on 2.13 trillion tokens on the Condor Galaxy, one of the world’s most powerful AI supercomputers.
In evaluation tests by the research team, Llama-3-Nanda-10B-Chat outperformed existing Hindi-centric and multilingual models across knowledge, reasoning, and misinformation benchmarks.
The 10B model incorporates over 100 safety categories and 50,000 adversarial prompt tests, designed to improve model safety and reduce risks of harmful or biased responses.
Users and developers can download code via Hugging Face and run NANDA locally under the terms of the Llama 3 licence.
The initiative could opens new commercial opportunities for Indian developers looking for a high-performance Hindi language model to use in the development of GenAI solutions.
The release of Llama-3-Nanda-10B-Chat also underscores the commitment of G42 group and MBZUAI to open-source.
LINKS
Llama-3-Nanda-10B-Chat code (Hugging Face)
Llama-3-Nanda-10B-Chat resources (Github)
Llama-3-Nanda-10B-Chat research paper (Github, PDF)
Find our more about NANDA and G42’s India initiative
G42 launches new Hindi LLM in Mumbai (Middle East AI News)
G42 to deploy 2GW data centre & supercomputer in India (Middle East AI News)