Reusing existing models

Model initialization controls whether a new draft starts with random weights or weights from a trained version of the same model.

Initialization modes

Mode	Source	Epoch setting	Use when
From scratch	None	Max epochs	You need an independent training run
Resume from checkpoint	A prior trained version of this model	Additional epochs	You want to continue the same model output
Transfer from another model	A prior trained version of this model	Max epochs	You want source weights but a newly trained version or output head

Only eligible trained versions of the current model appear as sources. Clone version (new model) copies settings into a separate model record, but it does not carry trained weights.

Configure initialization

1. Open a draft model version.
2. On Overview, expand Advanced Settings.
3. Open Model initialization.
4. Choose From scratch, Resume from checkpoint, or Transfer from another model.
5. For resume or transfer, select the source version and checkpoint selection.
6. Review the epoch count and optional random seed, then train.

For Best checkpoint selection, choose the type-specific metric that determines the best checkpoint. Most recent selects the latest available checkpoint instead.

Resume versus transfer

Use Resume from checkpoint when the model output remains compatible with the source. Resume adds Additional epochs and does not use the draft's Max epochs value.

Use Transfer from another model when you want to initialize a fresh training run from source weights. Classification and unmixing models also offer Head init:

• Mapped reuses compatible output-head weights for shared classes.
• Random initializes the output head again.

Evaluate candidate versions on Validation data, then evaluate the selected version on untouched Test data. Reusing weights can help, have no effect, or preserve unwanted bias from the source data.