Skip to main content
Synth’s supervised fine-tuning (SFT) loop has five steps. Each section below links to the deeper documentation when you are ready for details.

​
1. Publish a traceable task app

Wrap your environment with TaskAppConfig, enable tracing, and confirm the HTTP contract.
→ Read: Task App expectations

​
2. Collect rollouts

Use uvx synth-ai eval (or your own driver) to generate structured traces from the task app.
→ Read: Rollouts & evaluation

​
3. Filter to JSONL

Transform the trace database into SFT-ready dialogue records, applying any quality filters.
→ Read: Filtering traces

​
4. Launch training

Validate the JSONL, upload it, and start an SFT job with uvx synth-ai train --type sft.
→ Read: Training jobs

​
5. Ship and benchmark

Deploy the resulting fine_tuned_model, then re-run rollouts to verify improvements.
→ Read: Deploy with the CLI