Overview - Monte Platform

An environment is the package that a model is scored or trained in. Monte builds environments on NeMo Gym, so any Gym environment works unchanged.

Components

Component	Description	Example
Tasks	A JSONL file of task rows: the model’s input, its tools, and the fields the grader reads	A support ticket, plus the expected route, category, and facts
Server	A Gym resources server that runs during each attempt, with the tools and any state	`search_kb` and `escalate_to_human`
Grader	The server’s `verify` method, which returns a reward from 0 to 1	Checks the route, the category, and the facts in the reply

The same grader gives the eval score and the training reward. Rows without a grader are a dataset, not an environment.

Choose an Eval or a Training Environment

An environment holds one set of tasks, so it is an eval environment or a training environment, never both.

	Eval environment	Training environment
Purpose	Scores a model. Never trained on.	Gives a training run its tasks and reward
Used by	An experiment, as a held-out or development eval	`monte train --data <name>`
Add with	`monte env add --eval`	`monte env add --training`

Versions

monte env add makes an immutable version, such as support-escalation@2, with a fingerprint over the grader code, the task rows, and the configs. An experiment pins each eval version, and a run refuses to start when the fingerprint does not match, so a changed grader needs a new version and a new experiment.

Create and Add an Environment

monte env create support_escalation      # writes monte-gym/resources_servers/support_escalation/
monte env add --eval support_escalation --as support-escalation
monte env add --eval support_escalation_dev --as support-escalation-dev
monte env add --training support_escalation --as support-escalation-train

--eval takes an eval config under monte-gym/benchmarks/. --training takes a server whose config declares training rows. Tasks shows where each file goes. Environment Commands lists every flag.