Model Management
A single training run produces multiple progress saves (checkpoints) according to the configured save frequency. Model management is the function area for handling these artifacts.
Functions
| Function | Description |
|---|---|
| Share a specific checkpoint | Deployment generally requires only one save, so the complete set need not be transferred |
| Generate a download link | Packages the selected checkpoint as a downloadable file |
| Import a model | Uploads a model file, or receives one from another host via short code |
| Delete checkpoints | Progress saves occupy considerable disk space |
Retention of checkpoints
It is recommended that only the verified checkpoint be retained and the remainder deleted.
Before deletion, the checkpoint must be confirmed as viable on the inference page. The policy path on the inference page marks the loaded save as "model name / step count", which constitutes the basis for distinguishing between checkpoints.
One circumstance warrants attention: where loss begins to rise in the later stages of training — that is, where overfitting has occurred — the best-performing model is the save point preceding the rise rather than the final save produced. The version that actually performs best should therefore be confirmed before deletion.
Importing models
Two methods are provided:
First, file upload. Select or drag the model file into the upload area; the supported formats are displayed on screen. The platform performs a basic file validation prior to upload.
Second, short-code reception. For models shared from another host, enter the code in the input field and select "Receive"; the platform will download and extract the model into the model directory automatically. See Cross-Host Transfer.
Once imported, the model may be selected on the inference page via "Browse".
Deletion
Model management supports the deletion of an entire model or of an individual checkpoint. As with datasets, deletion cannot be reversed and the system requests confirmation.
Next: Cross-host transfer.