uvx synth-ai setup once so the CLI can load your SYNTH_API_KEY/ENVIRONMENT_API_KEY, then follow the PipelineRL loop below. Every stage is handled by uvx synth-ai—no SDK scripts or local servers required beyond your task app deployment.
- Publish a traceable task app – Wrap your environment with
TaskAppConfig, enable tracing, and confirm the HTTP contract (/info,/task_info,/rollout). → Task app expectations - Deploy the task app (Modal recommended) – Use
uvx synth-ai deploy --runtime modalso rollout actors can reach a stablehttps://<task>.modal.runendpoint. → Modal deployment - Author the RL config – Declare models, rollout settings, judges, and compute requirements in TOML. → Config reference
- Launch PipelineRL – Submit the job via
uvx synth-ai train --type rl --config ... --task-url <TASK_APP_URL>and stream progress. → Training guide - Evaluate + iterate – Use
uvx synth-ai eval(pointing at the same Modal URL) and the curated examples to benchmark new checkpoints. → Examples