Recover runs and manage pipeline versions

Continue attempts to resume eligible interrupted, failed, or ended work from the latest usable state instead of starting from the beginning.

When training completed but evaluation did not, an eligible Resume action evaluates the existing checkpoint without repeating training.

After completion and while idle, Run Again starts another run in the same pipeline lineage. Confirm the epoch prompt and configuration before submitting.

Clone creates a separate pipeline for a materially different experiment or configuration. It may consume additional workspace capacity.

Fine-tune creates a child pipeline from a completed trainable, non-classical model and a compatible dataset. Choose whether to keep the existing configuration and provide a focused instruction.

Choose a completed run version rather than assuming the newest visible result is the one used for a prior decision.

In the lower Results area, read the backend recorded for the selected version. Switching result tabs updates that label, and the version table shows backend context per version when it was recorded.

Review configuration, checkpoint, pinned backend, primary metric, detailed evaluation, and creation time for each version.

When sharing a result or deploying a model, record the pipeline and selected model version in your own change record.

Keep the pipeline private, share it with the organization, or make it global only when the workspace permits that scope and the content is suitable.

Pipeline collaborators can be Owner or Viewer. Use Viewer for inspection; reserve ownership for users responsible for lifecycle actions.

Choose whether the ZIP should include an eligible checkpoint and available progress artifacts. Treat the download as sensitive model material.

Store the snapshot securely and keep its pipeline-version context. Test import in an appropriate workspace before relying on it for recovery.