How GFT works

GFT improves a model by letting it work on tasks and scoring the result. This page explains the objects you meet in Studio, the CLI and the API.

Gyms#

A gym, also called an environment, packages three things: tasks, the actions a model may take, and a check that scores each attempt. Gyms come from three places:

• GFT gyms that GFT maintains.
• Private gyms that your organization publishes. Only your organization sees them.
• Marketplace gyms that partners publish and your workspace installs.

A gym is training-ready when it has training tasks, a reward profile and a runtime. It is evaluation-ready when it also has held-out tasks and a benchmark protocol. You choose a gym by its slug; GFT pins the exact version for each run.

Runs#

A training run improves a base model. Studio supports supervised fine-tuning (SFT) from examples, and reinforcement learning (RL) with GRPO on a gym. The CLI starts RL runs.

An evaluation scores a model on the gym's held-out tasks. The model never trains on these tasks, so the score shows how it handles new work.

Training runs and evaluations share one set of commands: list, show, watch and cancel.

Adapters and deployments#

Training produces a LoRA adapter: a small set of weights applied on top of the base model. You can download, evaluate or deploy it.

A deployment serves a model or adapter behind an authenticated inference endpoint. A new deployment must pass a canary check before it becomes active.

Workspaces, keys and budgets#

Everything belongs to an organization and one of its workspaces. API keys carry scopes that limit what they can do. Each paid job has a spending limit that you set before it starts, and is paid from your prepaid credit.

Next steps#

• Quickstart starts your first run.
• Create a gym turns your tasks into a gym.