Developing Malayalam LLMs will be the mandate of IBM’S GenAI Innovation Centre in Kochi

IBM's GenAI Innovation Centre in Kochi aims to enhance Malayalam Large Language Models using personalized data for productivity.

Developing Malayalam Large Language Models (LLMs) from its present generic state to more specifics facilitating training using personalised data remains part of the mandate of IBM’s proposed GenAI Innovation Centre in Kochi, according to Dinesh Nirmal, senior vice president, IBM.

Speaking to The Hindu on the sidelines of the ongoing International Conclave on GenAI here on Thursday, Mr. Nirmal said deliberations over further developing Malayalam LLMs were under way.

“The existing Malayalam LLM is just chat-based. The idea is to go into the next specific questions. What is crucial is, can I train that model using my data. Then I can ask my specific questions. If everyone can do that, just think about the productivity,” he said.

Mr. Nirmal said IBM’s proposed GenAI Innovation Centre in Kochi would be up and running this year itself for which the process was under way. He identified data and InstructLab as the two focal points of the proposed centre. IBM and Red Hat recently launched InstructLab to advance LLMs with open-source innovation.

“Data is very critical in GenAI. The critical thing is to ensure that the data is clean. InstructLab is about fine-tuning a model multiple times without creating newer versions of that model. The whole idea of InstructLab is democratisation of GenAI and help more and more people use it. We have even open-sourced it, and that work is being done out of Kochi,” he said.

On the need for an AI governance protocol, Mr. Nirmal drew attention to the formation of an AI alliance to help all stakeholders, including governments, start-ups, non-profit organisations, and universities, come along. “We had more than 8,000 different organisations sign up for it,” he said.

“If there is a force like that, then it is much more easy for us to ensure that the set of guardrails are there. The governments will also come up with a set of regulations that will enable a more methodical way of fine-tuning models since the data will become more creditable,” he added.