Lesson 4: Running a pretraining job, start to finish

Module 8 · Lesson 4 · About 50 minutes · Needs: Module 6's lab (train with checkpoints, resume and write_manifest); Lesson 1 (the lookup probe)

The situation

Sol 421 at Aurora Station. Causal language modelling won Lesson 2's race. Kofi Mensah now wants one real pretraining job on base, run the way large teams run theirs: "Plan it before you launch, watch it, show me it survives a crash, then hand me a report."

The station's CPU servers are shared, and a life-support job can take one without warning. The running example is the job itself, runs/m08-job.