Training methods
Choose SFT, preference learning, or RL only after evidence and evaluation.
Training methods
Training changes model weights; Harness and Skill refinement change the behavior around a model. Start with the smaller, reversible change when the evidence supports it.
| Method | Best fit | Requires |
|---|---|---|
| SFT | Clear input/output demonstrations | Curated demonstrations and a holdout evaluation |
| Preference training | Relative quality judgments | Comparable candidate outputs and reliable preference data |
| RL / GRPO | Measurable outcomes over multi-step rollouts | A bounded environment, reward/grader, budgets, and checkpoints |
Every method needs approved data, a candidate evaluation, a budget, and an explicit activation decision. Training is not implied by a Refiner proposal.
Next: Evaluation before choosing a training run.