The core loop: a closer look

The core loop #

For a weight recipe, the trainer does not run optimizer steps. It submits a bounded training job to a training component, such as Slime or Tinker. The component trains the job's data with its own training loop and saves checkpoints. Each saved checkpoint is a candidate. Serving loads only a snapshot that Reef published, and the next job starts from the newest saved checkpoint. A harness recipe uses its own backend in place of the training component: the backend prepares a candidate tree for each step. See Training components and Serving saved weights.

Learn more #

This page stays at the level of the loop. Each step it shows is specified in full elsewhere: