uvx synth-ai setup once so the CLI can load your SYNTH_API_KEY/ENVIRONMENT_API_KEY, then follow the loop below. Every stage relies on first-party CLI commands (no SDK scripts required).
- Publish a traceable task app – Wrap your environment with
TaskAppConfig, enable tracing, and confirm the HTTP contract. → Task app expectations - Collect rollouts – Use
uvx synth-ai eval(local or Modal via--url) to gather structured traces. → Rollouts & evaluation - Filter to JSONL – Convert trace databases into SFT-ready dialogues with
uvx synth-ai filter. → Filtering traces - Launch training – Validate/upload the JSONL and start a job via
uvx synth-ai train --type sft. → Training jobs - Ship + benchmark – Deploy the resulting
fine_tuned_model, re-run evals, and iterate. → Deployment
train/supervised-fine-tuning/ for detailed instructions, config schemas, and troubleshooting tips for each step.