Fine-tuning vs distillation
| Approach | Best for |
|---|---|
| Fine-tuning | Improving quality on a task where the base model is close but not good enough |
| Distillation | Preserving task quality while moving into a smaller, faster, cheaper student model |
The self-serve workflow
- Capture or import representative data in /observe/datasets-and-uploads.
- Define a baseline in /evaluate/overview.
- Launch a training run against paired training and eval datasets.
- Compare the resulting model against the baseline.
- Promote the winner into /deploy/overview.
What to optimize for
- Higher task accuracy on the prompts that matter to your product
- Lower cost by distilling into a smaller student model
- Lower latency for user-facing workflows
- Tighter output behavior for structured extraction, tagging, or classification workloads
Before you train
Do not skip the eval step. Training without a stable rubric and representative eval dataset makes it much harder to tell whether the new model is actually better.Next steps
Turn Eval Failures into a Training Run
Use a stable eval baseline and paired datasets to launch training the right way.
Create Datasets from Observed Traffic
Build the paired training and eval datasets required for self-serve training jobs.
Create datasets
Build the training and eval data you want to optimize against.
Launch and monitor training
Start a training job, inspect logs, and watch checkpoint evals.
Promote to deployment
Turn the completed training output into a production serving path.
Talk to an engineer
Meet with our team if you want help with dataset strategy, distillation, or rollout planning.