Overview

Synth’s supervised fine-tuning (SFT) pipeline turns high-quality dialogue traces into updated base models. The workflow mirrors what you would perform manually, but packaged so an agent (or CI job) can run it end-to-end.

CLI deep dives live under eval, filter, train, deploy, and serve; reach for those pages when you need every flag.

Lifecycle at a glance

1. Instrument a task app – expose your environment through TaskAppConfig, enable tracing, and record rollouts with the uvx synth-ai eval command.
   → See Task app requirements.
2. Filter traces into JSONL – convert the trace database into SFT-ready conversations, optionally filtering by score, split, or metadata.
   → See Filtering traces.
3. Validate and launch training – check schema compliance, upload the dataset, and submit the job via uvx synth-ai train --type sft.
   → See Training jobs.
4. Track progress and download checkpoints – monitor the job, capture the emitted fine_tuned_model id, and promote it to production.
   → See Deploying results.
5. Benchmark the new model – run the same rollouts (evaluation = rollouts) against the updated model to verify gains.
   → See Rollouts & evaluation.

Key commands

Action	Command
Run rollouts	uvx synth-ai eval --app-id <task-app> --model <base-model> --seeds 1-10
Filter traces into JSONL	uvx synth-ai filter --config path/to/filter_config.toml
Validate a dataset	uvx synth-ai train --type sft --dataset path/to/data.jsonl --config path/to/train.toml --no-poll (fails fast on invalid JSONL)
Submit training job	uvx synth-ai train --type sft --config path/to/train.toml --dataset path/to/data.jsonl
List fine-tuned models	curl -H "Authorization: Bearer $SYNTH_API_KEY" https://agent-learning.onrender.com/api/learning/models

Each sub-page dives into the specifics required to make an agent autonomous—from the TOML schema to the exact environment variables that the CLI expects.

Tip: the CLI auto-loads the .env written by uvx synth-ai setup. Pass --env-file only when you need to override the detected environment.