Getting started

Your first training run

Use the Studio web app to start a small reinforcement learning run on a hosted GPU. You need a Studio account, a training-ready gym in your workspace and prepaid credit. You can review the plan before committing any funds.

Start in the browser#

Open RL training in Studio and sign in. If your workspace needs credit, open Billing.
Choose a training-ready gym, then select a supported base model and a small number of training steps.
In Compute and payment, select an available GPU configuration. Review the estimated cost, compute reserve and gym spending limit. Confirm only when you are ready to launch.
Open Metrics and save the run ID. Check that episodes finish and produce rewards. When a checkpoint is available, evaluate it on held-out tasks.

For a private gym on your own Kubernetes cluster, use the cluster training guide. Your administrator retrieves the weights from the cluster's output volume.

Install the CLI#

The remaining steps are optional. If your workspace administrator has supplied the GFT client wheel, install it with Python 3.11 or later and uv. Set GFT_CLIENT_WHEEL to the path of the wheel you received.

bash

uv tool install "$GFT_CLIENT_WHEEL"

Connect to GFT Studio#

Point the CLI at the hosted API. Without this variable, the CLI calls a local development server at http://127.0.0.1:8787.

bash

export GFT_API_URL=https://studio.g-ftech.com

Create an API key on Settings → API Keys with the Read and Run scopes. Then sign in; the CLI asks for the key in a hidden prompt:

bash

gft login
gft doctor

gft doctor checks that the API is reachable and that your key works. See Authentication for scopes and automation.

Pick a gym#

List the gyms that are ready for training:

bash

gft env list --ready

The ID column is the gym's slug, such as shopify. Every command takes the slug; the CLI resolves the exact gym version for you.

Start a run#

Start a 100-step run with a $5 spending limit, and watch it until it finishes:

bash

gft train start shopify --model Qwen/Qwen3.5-9B --steps 100 --budget 5 \
  --follow --exit-status

The CLI asks you to confirm the cost, then prints the run ID and each status change. Stopping the watch with Ctrl+C does not cancel the run. Reconnect with gft run watch <run-id>.

If something fails#

What you see	What to check
`CONNECTION_ERROR`	Set `GFT_API_URL=https://studio.g-ftech.com` in the same terminal.
`UNAUTHORIZED`	Run `gft login` again with a valid key that has not expired.
`FORBIDDEN`	The key is missing a scope. Create a key with Read and Run.
`BUDGET_EXCEEDED`	Add credit in Billing or lower `--budget`.
`CAPACITY_REACHED`	Wait and try again, or start a smaller run.
`NOT_FOUND`	Copy the slug from `gft env list --ready`.

Every error includes a request ID. See Errors and exit codes.

Next steps#

How GFT works explains gyms, runs, adapters and budgets.
Evaluate a model measures your result on held-out tasks.
CLI reference lists every command.