Private and marketplace gyms

The Training picker lists every active gym in your workspace: private gyms that your organization published, and marketplace gyms that your workspace installed.

Train with a workspace gym#

1. Open Training and pick the gym.
2. Select a model.
3. Set limits for GPU time and gym usage.
4. Review the combined reserve and start.

A private gym can also train on your own cluster with a matching cluster profile. See Connect a gym to a cluster.

Installed marketplace gyms run on GFT Cloud through the metered workspace gateway. They cannot run on cluster profiles. Check installation status on the Marketplace.

Judge a marketplace gym before you install it#

Open a gym from the Marketplace to see its page. It shows one example task, how the reward is computed, the task counts, and the price of a full episode. Two sources are labeled apart:

• Measured by GFT from the qualified package: task counts, steps per episode, reward parts and their weights, and the qualification checks.
• Reported by the author: the example task, the shortcuts tried against the reward, and model results. GFT does not reproduce these.

Train from the CLI#

gft train start creates a plan for a configured training worker. An accepted plan does not prove that a trainer is running, so confirm that a worker is available with your operator before you reserve a budget. See Train with RL.