Skip to main content
Every synth-ai train flow is driven by a TOML config that maps to a Pydantic schema under synth_ai/train/configs/. The CLI loads your file, validates it, and turns it into the payload sent to the backend. Use this overview to choose the right schema and find the detailed references.

​
Schema Cheat Sheet

FlowDocSchema Class
Reinforcement LearningRL Configsynth_ai/train/configs/rl.py (RLConfig)
Supervised Fine TuningSFT Configsynth_ai/train/configs/sft.py (SFTConfig)
Prompt Learning (GEPA/MIPRO)Prompt Learning Configsynth_ai/train/configs/prompt_learning.py (PromptLearningConfig)

​
Shared Guidance

  1. Match the schema exactly: Field names/types in your TOML must align with the schema attributes; the CLI calls *.from_path() directly.
  2. Keep secrets in .env: Task URLs belong in TOML, but API keys should be stored in environment files loaded via uvx synth-ai setup or --env-file.
  3. Validate early: Run uvx synth-ai train --type <flow> --config <file> --no-poll to surface schema errors without waiting for training to finish.
  4. Document decisions: TOML comments (# ...) are ignored by the parser—use them to explain dataset locations, model choices, or intended usage.
  5. Version control configs: Commit configs alongside task app changes; reproducibility depends on syncing code + config.
Head to the flow-specific pages above for required sections, optional fields, and sample snippets.