How to train LLMs to beat frontier models

Real phone conversations. A small model. A frontier-model comparison.

See Phonely’s approach to turning messy production data into a training set, fine-tuning on AWS GPUs, and evaluating whether the model actually gets better.

Registration

1. RSVP on this Luma page.
2. Complete your AWS Builder Center registration for venue entry.

Both steps are required. A Luma RSVP does not automatically register you with AWS or grant entry to the AWS Builder Loft. Everyone attending, including speakers and hosts, must register with AWS.

Event details

Join Will Bodewes and Amit Kiran Rege from Phonely for an evening of fine-tuning a small model to hold a conversation. We’ll start with a set of real, scrubbed phone conversations, some of which contain bad responses, and train a small model to respond the way a person would.

We’ll walk through Phonely’s approach to cleaning the dataset, running the fine-tune on AWS GPUs, and building an eval that compares the base model, the fine-tuned model, and a stronger frontier model.

In the course of this event, you will learn how to turn messy production data into a training set, how to tell whether a fine-tuned model is actually better, and how to read its failures to decide what to change next.

Speakers

• Will Bodewes, CEO, Phonely
• Amit Kiran Rege, ML Researcher, Phonely

Who it’s for

Engineers, ML and AI practitioners, and technical founders. Intermediate level.

Food and soft drinks provided.

Entry requirements

All guests must be 18+ and show a valid government-issued physical photo ID at check-in. Digital IDs are not accepted. Scooters and bikes are not permitted in Amazon buildings. No on-site parking is available.