Run, monitor, and compare pipelines

Start the reviewed pipeline a single time. A preparation, planning, pending, or queued state means work was accepted; do not submit a duplicate.

Wait while samples, splits, and task inputs are prepared. Read the visible message history when preparation reports a problem.

Use the queue state, progress indicator, current message, and message history to distinguish waiting from active training or evaluation.

If the live display disconnects, refresh the page and use the persisted status and history before deciding whether intervention is needed.

Interrupt requests a controlled stop. Wait through Interrupting until the pipeline reaches Interrupted or another terminal state.

Compare required memory, available or free memory, the current batch size, and the recommended batch size shown by the dialog.

Use the suggested lower batch size when offered, then save the configuration.

Run the preflight check again after changing the configuration. Retry the pipeline only when the check permits it.

Start from a completed pipeline with results.

Search for pipelines that evaluate the same task and comparable target. Incompatible or incomplete pipelines are not useful comparison candidates.

Check dataset, split, model version, metric definition, threshold or operating point, and result status before interpreting differences.

Use the comparison as navigation, then inspect each pipeline result and model report for the evidence behind the headline metric.