Skip to main content

Train

synth-ai train Interactive launcher for RL / SFT / Prompt Learning / GraphGen / Context Learning jobs.

Arguments

  • CFG_PATH (optional)

Options

OptionTypeDefaultDescription
--envPATH-.env file(s) to preload (skips selection prompt)
--task-urlTEXT-Override task app base URL (RL only)
--datasetPATH-Override dataset JSONL path (SFT)
--modelTEXT-Override model identifier
--allow-experimentalflag-Allow experimental models (overrides SDK_EXPERIMENTAL env)
--no-allow-experimentalflagfalseDisallow experimental models (overrides SDK_EXPERIMENTAL env)
--idempotencyTEXT-Idempotency-Key header for job creation
--pollflagtruePoll job status until terminal state
--poll-timeoutFLOAT3600.0Maximum seconds to poll before timing out
--poll-intervalFLOAT5.0Seconds between poll attempts
--stream-formatChoice(cli, chart)cliStreaming output style (cli = line updates, chart = live loss panel)
--examplesINTEGER-Limit SFT training to the first N examples
--backendTEXT-Backend base URL (e.g., http://localhost:8000). Overrides BACKEND_BASE_URL env var.
--local-backendflag-Use local backend (localhost:8000). Overrides TOML [display].local_backend
--tuiflag-Enable live TUI dashboard. Overrides TOML [display].tui
--show-curveflag-Show optimization curve at end. Overrides TOML [display].show_curve
--verbose-summaryflag-Show detailed final summary. Overrides TOML [display].verbose_summary
--typeChoice(prompt, rl, sft, graphgen, adas, context_learning)-Explicitly set training type. Required for GraphGen (uses JSON datasets). ‘adas’ is a legacy alias.
--rollout-budgetINTEGER-Rollout budget for GraphGen optimization (default: 100)
--proposer-effortChoice(low, medium, high)-Proposer effort level for GraphGen (default: medium)