Overview - Monte Platform

Train an open-weight model on a team’s own tasks. A training run takes a training config, training data, and a starting point, and saves checkpoints as it trains. Every checkpoint can be scored on the experiment’s evals and deployed.

monte train support-triage \
  --config triage-sft \
  --data support-replies \
  --steps 200 \
  --from-base

Flag	Description
`--config`	The training config: model, method, LoRA, batch size, learning rate
`--data`	The training data: a dataset or a training environment
`--steps`	The number of steps in this run
`--from-base` or `--parent <run>/<step>`	The starting point: the base model, or a checkpoint

The run saves a checkpoint every 50 steps and at its last step, so this run ends with train-07/200. A run starts only after every eval in the experiment has a baseline.

Choose a Method

The training config sets the method. The method sets the kind of training data.

Method	Signal	Training data	Use when
SFT	Example responses	Dataset, `--kind sft`	The target output can be written down
DPO	Pairs of a preferred and a rejected response	Dataset, `--kind preference`	Reviewers can pick the better of two outputs
GRPO	The grader’s score on the model’s own samples	Training environment	A grader, in code or with an open-weight judge, can score an output
Distillation	The teacher’s next-token distributions on the student’s own samples	Training environment	A larger model already does the task well

Choose Full Fine-Tune or LoRA

A lora block in the training config trains low-rank adapters, and the base weights do not change. Without the block, the run trains every weight.

	Full fine-tune	LoRA
Checkpoint	A full model	An adapter
Serving	One deployment for each checkpoint	Many adapters on one base deployment. See Serving.
Use when	The task needs the largest change to the model	Several variants of one model, or a fast first run

Continue From a Checkpoint

--parent train-07/200 starts a new run from that checkpoint with any config and data, and monte train resume continues a run for more steps. A LoRA run with --parent on a LoRA checkpoint continues the same adapter, so it serves on the same base. Several runs from one parent give several adapters on one base.