Training and evaluation
Train with reinforcement learning
Start a GRPO training run on a gym from the CLI, then follow, inspect or cancel it.
You need the CLI signed in with the run scope, and a training-ready gym. See Quickstart.
Start a run#
gft train start shopify \
--model Qwen/Qwen3.5-9B \
--steps 100 \
--group-size 4 \
--episode-steps 8 \
--budget 5 \
--follow --exit-statusThe CLI resolves the active version of the gym, pins it in a plan, and starts one run. Network retries inside the CLI never start a second run.
| Option | Default | Meaning |
|---|---|---|
--model, -m | required | Base model name or immutable revision. |
--steps | 100 | Number of optimizer updates. |
--group-size | 4 | Rollouts GRPO compares for each prompt, from 2 to 16. |
--episode-steps | 8 | Maximum tool actions inside one rollout. |
--budget | 5 | Maximum spend in USD. |
--follow, -f | off | Watch the run until it finishes. |
--exit-status | off | With --follow, exit with an error unless the run completes. |
--yes, -y | off | Skip the cost confirmation. |
Follow and manage runs#
gft run list --kind training
gft run show <run-id>
gft run watch <run-id> --exit-status
gft run cancel <run-id>The gft run commands work for training runs and evaluations. Cancelling twice is safe.
Where training runs#
/v1/runs accepts a run, checks the budget and opens gym sessions. It does not start a GPU by itself: the server must connect accepted runs to a TRL training worker. Provider settings stay on the server, so the CLI never needs a second URL.