stateset-agents --help and each subcommand’s --help from the reviewed environment.
These examples follow the source scope of the latest checkout.
Fine-tuning control plane
These commands accept
--workspace to share the same retained runs with the TUI. See the
complete prepare/review/launch workflow.
Recovery and evaluation
Recovery preserves the original budget. Validation selects within a matching protocol;
fresh test uses separate cases. Neither command completion nor checkpoint recommendation
promotes a model. See recovery and evaluation.
Data and industry workflows
industry train --dry-run is a plan; real execution needs configured CUDA and
training dependencies. Keep validation and test rows out of training inputs.
Remote compute and inference
The studio requires
--allow-paid for remote launches. RunPod’s quoted pod-lifetime cap
excludes ancillary costs; Modal/River use explicit provider-billed policy without an enforced
dollar ceiling. An SDK cleanup acknowledgement does not prove terminal provider state or
that billing stopped. Inspect original handles and the provider account before replacing an
uncertain run. See River training.
MCP and HTTP
mcp --transport stdio exposes curation and free previews; it cannot launch training.
The newer authenticated /finetune operator HTTP routes inspect retained work and request
local stops. Their shared JSON semantics are documented in
agent fine-tuning controls.