Checkpoints

Training produces two kinds of checkpoint, with different contents and uses.

Type	Path format	Contents	Use
State	tinker://<session>/weights/<name>-state	Model weights + optimizer state	Resume training (continuation)
Sampler	tinker://<session>/sampler_weights/<name>-sampler	Model weights only	Inference / evaluation

You find both paths in the Training page's version records and the Recent training artifacts list.

Resume training (incremental continuation)

The training page supports resuming from a previous version's state_path.

import mint

sc = mint.ServiceClient()

# Restore weights only (optimizer reset — useful when switching datasets)
training_client = sc.create_training_client_from_state(state_path).result()

# Restore weights + optimizer (true continuation, preserves Adam momentum)
training_client = sc.create_training_client_from_state_with_optimizer(state_path).result()

In the MindForge training panel, set --resume-state to the previous version's state_path, and add --resume-with-optimizer for a full continuation that keeps the Adam optimizer state.

From checkpoint to inference

A Sampler checkpoint is already inference-ready — pass its path as the model field. See Using a trained model for the three inference methods.