Getting started
GFT Studio documentation
GFT Studio trains, evaluates and serves language models on tasks from your domain. A gym supplies the tasks and checks the model's work. Studio uses those checks to train a model, measure it on tasks it has not seen, and deploy the result behind an inference endpoint.
Use the Studio web app for guided workflows, or the gft CLI, Python SDK and REST API to automate the same steps. Paid work always shows a cost review or asks for a spending limit before it starts.
Start in the browserChoose a gym, review the budget and start your first training run.Use StudioFine-tune, evaluate and deploy from the web app.Bring your own gymPackage your tasks and scoring as a private gym.
What you can do#
| Task | In Studio | With the CLI |
|---|---|---|
| Fine-tune on examples (SFT) | Fine-tune | — |
| Train with rewards (RL) | RL training | gft train start |
| Evaluate on held-out tasks | Benchmarks | gft eval start |
| Publish a private gym | Private gyms | gft env add, gft env push |
| Deploy and chat with a model | Models, Playground | gft deploy, gft chat |
| Run your own container on a GPU | Containers | gft container run |
Find your path#
- New to GFT: Quickstart → How GFT works.
- Working in the web app: Use Studio → Run a result locally.
- Bringing your own tasks: Create a gym → Publish with the CLI → Connect a gym to a cluster.
- Building a gym for customers? Publish a marketplace gym.
- Training and measuring: Train with RL → Evaluate a model.
- Serving a model: Deploy an adapter → Chat and benchmark.
- Running your own code on GPUs: Run your own container.
- Automating: CLI reference · REST API · Errors and exit codes.