Train
Stage: Build · What it does: launch a training run from your explored data and ship the resulting model to Ship.
Why it matters
- Reproducible. The same run re-executes on new data instead of "someone re-ran the notebook."
- Lineage for free. Runs launched through the platform record what they read and write.
- Progress you can see. Runs stream metrics; interrupted runs resume from checkpoints.
How to use
- In Build, launch NN Training from the Launch a run actions.
- Define the source input — e.g. a pretraining corpus like FineWeb, or a file/glob from your connected sources.
- Configure the training steps.
- See the cost estimate, optionally tick GPU smoke, and launch.
- Track status in the run list; the completed model appears in Ship's deploy dropdown.
CLI users: the guided flow maps to a CLI command — the console is a visual front-end for the same job.
Real-world examples
🏦 Financial services — nightly credit-risk refresh
A bank retrains a scoring model every night before market open. A GPU smoke run before the full retrain catches a malformed data day at pennies instead of hours.
🏭 Manufacturing — morning quality model
An IIoT team retrains a quality model each morning over the previous day's sensor telemetry. An interrupted run resumes from its checkpoint, so a 5-minute network blip doesn't restart 40 minutes of processing.
Gotchas
- Use globs for rolling data so the same run consumes new partitions automatically.
- Runs have cost estimates and an optional GPU smoke preflight — use both; they're cheap insurance.
- Stopped/cancelled runs clean up their scheduler and partial state (no zombie jobs).