Setting up our research environment

Most ML research assumes access to a cluster. We do not have one. Our primary machine is an M1 MacBook, and we are building our research workflow around that constraint.

This post covers our setup: Python environment management, experiment tracking, how we handle model checkpoints with limited disk space, and the tools we use to keep experiments reproducible when compute is scarce.

Constraints force good habits. When you cannot afford to waste a training run, you plan more carefully.