GFT Studio
Run GRPO on your GPUs with GFT Studio.
GFT Studio trains open models on tasks you can grade. Use a training-ready gym on GFT-hosted GPUs, or run a private gym on your own Kubernetes cluster. Your choice determines where the training results live.
What do I need for a first run?
For a hosted run, start with a Studio workspace, prepaid credit and a training-ready gym. GFT supplies the GPUs. For a cluster run, your administrator needs a connected Kubernetes cluster, trainer and gym images, and storage for inputs and outputs.
From a task to an adapter
- Choose a task and a baseline. Browse the gym catalog and reserve tasks for evaluation before training. If you plan to use your own cluster, choose a private workspace gym. Marketplace gyms run on hosted compute.
- Launch a short run. For hosted training, choose a gym, model and GPU configuration, then review the budget. For your own cluster, open Settings → Kubernetes clusters. Your administrator binds the private gym version to a local training profile. Open “Train with this gym,” select that profile, review the limits and confirm.
- Evaluate a saved checkpoint. Compare the adapter and its original base model on the same held-out tasks. Track correctness separately from cost or speed. See how to compare checkpoints.
- Find the weights. For a hosted LoRA run, download the adapter from an exportable checkpoint. On your own cluster, Studio shows the run status and optional metrics. Your operator retrieves the weights from the output volume under the run ID. Keep the base model revision with the weights for inference.
A worked example with weights and results
In our 27B Triton kernel study, the model received a PyTorch operation and wrote a GPU kernel. The verifier checked that it compiled, returned the right values and ran fast enough. The study links the released LoRA adapter, training data and per-run results, and explains where training helped and where it stalled.
Explore the GPU kernel environment or discuss a gym for your own workflow.