Jillian Bommarito Discusses KL3M on the LawNext Podcast

273 Ventures Chief Risk Officer Jillian Bommarito was the featured guest on LawNext, the legal innovation podcast hosted by veteran legal journalist Bob Ambrogi.

The First LLM Built From Scratch for Legal

The episode focused on KL3M (the Kelvin Legal Large Language Model), which stands out in legal AI for two reasons. First, it is the first LLM built entirely from scratch specifically for the legal market, rather than fine-tuned from an existing general-purpose model. Second, it is the first LLM in any domain to be trained entirely on clean, legally permissible data and to be certified as such by the organization Fairly Trained.

Clean Data, Clean Models

Jillian discussed the process of building KL3M’s training dataset and the deliberate decision to use only lawfully obtained, low-toxicity data. She explained how she oversaw the Fairly Trained certification process, which involved an independent third-party audit of the training data and practices used to build the model.

As Ambrogi noted, regular LawNext listeners may remember his earlier interview with 273 Ventures CEO Michael Bommarito and Chief Science Officer Daniel Katz, who appeared on the show after conducting the first experiment in having GPT take the bar exam. All three, along with Katz’s wife Jessica Katz, previously co-founded LexPredict, which was acquired by Elevate in 2018.

Why Domain-Specific Models Matter

The conversation also addressed why a law firm or legal department would choose a purpose-built legal model like KL3M over commercially available general-purpose alternatives. Jillian explained the practical advantages: lower toxicity, better performance on legal language, and the ability to provide customers with representations and warranties about the training data, something that general-purpose model providers cannot offer.

Listen to the Episode

The full episode is available on LawNext and wherever you listen to podcasts.