Checkpoints¶
Preserve training artifacts and choose the right version for serving, evaluation, or continued training.
A checkpoint preserves a training artifact at a particular stage. Mint Recursive groups these artifacts into model versions associated with training history, parent versions, and deployments.
Versions, checkpoints, and deployments¶
| Object | Purpose |
|---|---|
| Model version | Select an artifact in My models and trace its training lineage |
| Sampler checkpoint | Weights used for inference |
| Training state | State for continuing training; optimizer restoration depends on the resume method |
| Deployment | A running service and its calling configuration |
Some backends expose sampler_path and state_path references. These are artifact references, not the Model ID shown on the API access page.
Choose a version¶
After training, inspect linked checkpoints in My training (我的训练) or trained versions in My models (我的模型). Select using the training record, dataset version, and evaluation results rather than recency alone.
Use Deploy (部署) to serve a version and Evaluate (评估) to test it. If a required artifact is missing, check that the training run completed and its output was registered.
Continue training¶
Choose Continue training (续训) for a version in My models. The new run inherits the base model, and you select its data and configuration. A complete resume requires recoverable training state; inference weights alone do not imply that optimizer state can be restored.
Retain parent versions and run history to assess the effect of each change. Evaluate old and new versions on a consistent test set before serving the new one.
Related: Training · Deployment · Evaluation