How to train your own LLM

It takes an extraordinary amount of effort to train a large language model. Until now, that is.

05 July 2024

Matt Hicks, Red Hat

While generative AI continues its seemingly unstoppable march into every IT service on the planet, it’s still early days for enterprise adoption. IBM reckons that 38% of IT professionals at enterprises report that their company is actively implementing GenAI and another 42% are exploring it. Large language models, upon which GenAI services are based, aren’t open source, or they come with a custom licence. While there are some open models, among them Llama from Meta, and Mistral, there’s little visibility into their training data, and there was no way that communities could contribute to LLM models. Red Hat and IBM think they may have found the answer. In May, Red Hat announced the developer preview of Red Hat Enterprise Linux AI (RHEL AI), on which firms can develop and test their open source Granite GenAI models. RHEL AI is based on the InstructLab open source project and uses the Granite large language models from IBM Research. InstructLab, in turn, is based on what is called LAB (Large-scale Alignment for chatBots) methodology, which is a new way of training LLMs and reduces reliance on expensive human annotators and proprietary models like GPT-4. It will come as a bootable RHEL image.

The announcement was front and centre at Red Hat’s annual summit in Denver, Colorado, in May, and CEO Matt Hicks said that InstructLab allows anyone – even this reporter – to contribute to and train LLMs. Hicks said while there has been much progress in the open-source ecosystem, the ability to contribute to an LLM had yet to be solved.

ITWeb Premium

Get 3 months of unlimited access
No credit card. No obligation.

Already a subscriber Log in