Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.arkor.ai/llms.txt

Use this file to discover all available pages before exploring further.

​
Trainer control

The Trainer returned by createTrainer has three methods: start, wait, cancel. arkor start and Studio’s Run training button call start() followed by wait() for you. You only call them yourself when wiring training into your own code (a server, a script, a custom CLI). 
const { jobId } = await trainer.start();
const { artifacts } = await trainer.wait();
  • start() submits the job and resolves once the backend accepts it.
  • wait() opens the SSE event stream, dispatches your callbacks, and resolves when the run finishes.
  • cancel() asks the backend to stop the run.
If you only need callbacks, you do not need to handle start / wait directly: arkor start does it. Reach for these methods when you are running outside the CLI.

​
Stopping a wait early

Pass an abortSignal to createTrainer if you need a way to stop your local wait() loop without waiting for the backend: 
const controller = new AbortController();

const trainer = createTrainer({
  /* ... */
  abortSignal: controller.signal,
});

// from somewhere else:
controller.abort();
abortSignal only stops your local loop. To also stop the GPU on the backend, call await trainer.cancel() afterwards.

​
Reference

For the exact reject vs resolve behavior, the SSE reconnect rules (initialReconnectDelayMs, maxReconnectDelayMs, maxReconnectAttempts), the idempotency guarantee on start(), and the full two-process pattern with SIGINT handling, see the Trainer control reference.