Rubrik Inc. today announced plans to acquire Predibase Inc., a startup that develops software for fine-tuning large language models.

CNBC reported that the deal is worth $100 million to $500 million. That represents a major exit for Predibase’s investors, who had provided the company with about $28 million in funding prior to the sale.

NYSE-listed Rubrik provides a data management platform used by more than 6,100 organizations. Many on the platform’s features focus on helping administrations avoid data loss. Rubrik can create backups of records stored in cloud environments, on-premises arrays and other sources, as well as restore them if an outage occurs. A recovery simulation tool allows administrators to test that their data protection workflows function as intended.

Another set of features in Rubrik’s platform promises to help companies secure their data. The software automatically detects files that contain sensitive records such as payment card details. It allows administrators to define policies that regulate which worker can access which record and how.

According to Rubrik, the data that enterprises manage using its platform is useful for artificial intelligence projects. The technology that Predibase brings to the table will enable the company to simplify such AI projects for its customers. “Together, we can equip AI teams with the tools and building blocks needed to quickly bring enterprise-grade AI applications to production,” Rubrik co-founder and Chief Executive Officer Bipul Sinha (pictured) wrote in a blog post today.

Predibase’s namesake platform enables developers to fine-tune open-source LLMs. Fine-tuning is a process that increases the quality of an LLM’s output by putting it through additional training. This training is done with examples of tasks highly similar to the ones the model will perform in production.

The company uses a technique called LoRA to streamline the fine-tuning process. It works by extending the LLM being customized with a small number of additional parameters, or configuration settings. This approach is considerably faster than the standard way of fine-tuning an LLM, which involves modifying all its existing parameters.

Predibase’s platform also supports a second fine-tuning approach dubbed Turbo LoRA. It combines LoRA with another machine learning method known as speculative decoding. The latter technology allows LLMs to generate multiple output tokens at once, which is faster than the usual approach of outputting tokens one at a time. Predibase says Turbo LoRA triples LLMs’ inference performance in some cases.

After fine-tuning an AI model, the company’s platform can deploy it to production using a technology called LoRAX. It’s an open-source tool that speeds up LLM inference by applying various performance optimizations. According to the company, LoRAX allows developers to run up to thousands of AI models using a single graphics card.

“Their tools allow customized models to run and scale — economically and quickly,” Sinha wrote today.

Predibase will operate as a standalone business unit of Rubrik after the acquisition closes. Its platform will remain compatible with the third-party data management platforms that it currently supports. 

Photo: Rubrik