Skip to main content
These commands train a model and manage the checkpoints that the runs save. Training Overview compares the four methods.

monte train

Start a training run in an experiment. The run saves a checkpoint every checkpoint_every steps of the recipe (default 50) and at its last step. Steps count from the base model.

monte train extend

Add steps to a finished training run, recorded as a new run that points to the original. The new run takes the parent’s recipe, training data, weights, optimizer state, and data position, so the curve continues from the last step.
To change the recipe or the data, use monte train --parent instead. That run takes the weights only.

monte checkpoint list

List the checkpoints of an experiment, newest first.
local means the weights are on this machine, and stored means they are in artifact storage. The kind in brackets is servable, resume, or both.

monte checkpoint show

Show one checkpoint in full: the run, the base model, the files and their sizes, the eval scores, and the deployments that serve it.

monte checkpoint push

Upload the weights of a checkpoint from this machine to artifact storage, so any GPU instance can evaluate, resume, or deploy it. --push on monte train does the same at the end of a run.

monte checkpoint download

Download the weights of a checkpoint from artifact storage. With no --dest, the files go under the data root, where monte eval and monte train find them. A servable folder loads with AutoModelForCausalLM.from_pretrained, or with PEFT for an adapter.

monte checkpoint delete

Remove the weights of a checkpoint from artifact storage and from the GPU instances. The run and its scores stay in run history.
A delete with --kind servable or all cannot be undone. The checkpoint can no longer be evaluated or deployed.