Checkpoints¶
A checkpoint is a saved training artifact for one experiment at a given step. Manage them from the Experiments tab, not Results.
Finding checkpoints¶
Expand an experiment row. The expanded area shows a run summary chart and a Checkpoints table.
Column |
Meaning |
|---|---|
Steps |
Training step when the checkpoint was written |
Training Score |
Score metric at that step when metrics exist |
Evaluation Score |
Fitness metric at that step when metrics exist |
Size |
Stored size, shown in MB in the table |
Actions |
Deploy and delete controls |
If metrics are not available yet, score cells show N/A.
Actions¶
Deploy checkpoint: opens the create-agent flow for that step. Pick a name and confirm; Arena opens Agents → Advanced Training when the experiment used a dataset, or Classic RL when it used a gym environment only (even on an Advanced Training project). Use Connect there to turn on HTTP inference.
Delete checkpoint: confirmation dialog names the step; confirm with Delete
Resume training on the experiment row opens a dialog that lists checkpoints for the same experiment, with the same score columns when metrics exist.
Pipelines and agents¶
Pipeline Experiment stages after the first can start from Best checkpoint or Final checkpoint of the previous stage. The first stage can Start from scratch or Use saved Agent. Deployment stages choose Best checkpoint or Final checkpoint from the prior experiment stage. See Pipelines.
LatentPPO stages note that the previous checkpoint becomes the decoder.
Experiment-stage checkbox: Auto-add checkpoint to Agents when complete.