Tutorial

The in-product walkthrough: four concepts, then five steps to a first run.

An onboarding page inside the console, for someone who has an account and has not run anything yet. It covers the same ground as the Quickstart, with the deployment’s own values filled in — its profile names, its enabled methods — rather than examples.

Four concepts first

Term	What it is
Project	One repository. Named by name in starforge.yaml, and what groups runs in the console
Experiment	One directory under experiments/. A reproducible training definition: config, lockfile, notes
Dataset	A versioned, checksummed platform asset. You declare a reference; the platform resolves it inside the job
Job	One submission. The server injects resources, secrets and data, then schedules it

The distinction people trip on is experiment versus job. An experiment is a definition you edit; a job is one execution of it. Submitting twice gives you two jobs and two runs from one experiment, which is exactly what you want when comparing seeds.

Five steps to a first run

The tutorial walks the same path as the Quickstart: create a project, choose a framework, point at data, validate, submit. Where it adds something is the framework comparison — the same experiment expressed against NeMo-RL, verl, TRL and OpenRLHF, so you can see what actually differs between them before committing to one.

Where to go after it

Quickstart

The same path in documentation form, with the confirmation checklist.

Submitting in depth

Every flag on the submit path.