Your dataset. Your next model.
Bring a dataset, choose a compatible model profile, and define the run. Prepare a supervised fine-tuning pilot—or save an RL configuration for later.
Open Training PRIVATE SFT PILOT / RL DRAFTS- Input
- Versioned JSONL dataset
- Profile
- Pinned compatible model
- Limits
- Steps and duration
- Output
- Model artifacts + run record
Illustrative SFT deliverable. No run is active and no result is claimed.
Keep the run and its outputs together.
An approved SFT pilot links its dataset, execution profile and run record. The selected profile defines the model artifacts. An RL draft is a configuration—not a trained checkpoint.
From configuration to next step.
- 01
Bring your dataset
Select a validated, compatible dataset from your private Datasets library, or upload a JSONL file in the format required by the selected SFT profile.
- 02
Define the run
Choose a compatible profile, name the project and set step and duration limits. RL configurations can be saved as drafts.
- 03
Review before execution
Pilot access, an active execution profile and explicit confirmation are required for SFT. Inspect the run record when an approved run completes.
Built around your workload.
Pilot scope and any execution quote are reviewed before a run is confirmed. Saving a configuration does not reserve compute or create a billable training run.
- Supervised fine-tuning
- CPU SFT for approved pilot teams, using the model, dataset format and limits exposed by a compatible execution profile.
- Reinforcement learning
- Draft-only DAPO configurations: model, environment selection, project objective, maximum steps and duration.
- Environment selection
- Code, math and logic in the reference RL catalog. A saved combination is not a qualified training runtime.
SFT by approval. RL in draft mode.
CPU SFT requires approved pilot access and a compatible active execution profile. GPU training and custom RL execution are not publicly available. Saving an RL draft does not launch training.