Prepare training data
Separate demonstrations, reward tasks and held-out evaluation before preparing a run.
Start from a versioned Dataset. Retain its source, selected tasks and exact release. Choose the learning signal before preparing compute: SFT needs reviewed desired answers; GRPO needs tasks whose fresh attempts can be scored by an executable reward.
Keep the roles separate
| Component | Receives |
|---|---|
| Model making an attempt | Task instructions, permitted context, tools and environment observations |
| Grader | The response or actions, scoring rules and permitted reference information |
| RL optimizer | Generated attempts and their rewards |
| SFT optimizer | Reviewed training context and assistant targets |
| Evaluation report | Separate held-out results for comparison and review |
A corrected answer needs explicit review and admission before it becomes an SFT target. Human labels can improve a grader without being inserted into the model's prompt. Keep private references and judge examples with their intended owner. Do not turn frozen evaluation tasks into training examples.
Prepare a reviewed release
- Build or import the Dataset and inspect its provenance, input/output schema, references and training/evaluation split.
- Choose the learning signal and review the actual targets or reward checks. A successful follow-up tool call is a recovery candidate, not automatic training permission.
- Validate the selected graders and establish a baseline on separate tasks. Resolve execution and calibration blockers before approving training.
- Select the exact published release in the experiment. Review the compatible base, method, runtime, prepared inputs, price and maximum spend before Start.
| Preparation task | Existing workflow |
|---|---|
| Build or import tasks | Datasets |
| Label useful answers and admit examples | Review and label tasks |
| Measure the starting behavior | Experiments |
| Qualify a judgment-based reward | Improve an LLM judge |
| Use a deterministic rule or native verifier | Code verifiers |
| Update changed scoring criteria | Grader versions |
Start from a benchmark
A benchmark Profile can provide tasks, environment instructions and native scoring. Inspect the source, selected cases, grader and execution requirements before running it. Reuse its native scoring rather than recreating it.
For example, τ Retail's verifier can check the resulting database state. A conversation that sounds successful can still fail that outcome check. A local verifier check does not qualify execution in another hosted environment.
Choose the execution guide
Use SFT with LoRA for the reviewed-demonstration flow, or GRPO for reward-based setup and continuation. Model modifications covers other methods and their distinct input roles.
For recurring checks, submit only the selected published training tasks to the Model's review queue. A Dataset save does not admit every row or schedule compute. Follow Continual learning setup for source binding, schedules and limits.
Compatibility
A published Dataset or successful local verifier check does not qualify a hosted training runtime.