Callbacks

Watch a training run from your code.

The SDK fires callbacks at five moments while a run streams: started, log, checkpoint, completed, failed. They let you forward metrics, evaluate intermediate checkpoints, or trigger your own notifications without leaving TypeScript.

import { createTrainer } from "arkor";

const trainer = createTrainer({
  name: "support-bot-v1",
  model: "unsloth/gemma-4-E4B-it",
  dataset: { type: "huggingface", name: "arkorlab/triage-demo" },
  callbacks: {
    onStarted: ({ job }) => console.log("started", job.id),
    onLog: ({ step, loss }) => console.log("step", step, "loss", loss),
    onCheckpoint: async ({ step, infer }) => {
      const r = await infer({ messages: [{ role: "user", content: "Hi" }] });
      console.log("checkpoint", step, await r.text());
    },
    onCompleted: ({ job }) => console.log("done", job.id),
    onFailed: ({ error }) => console.error("failed", error),
  },
});

await trainer.start();
await trainer.wait();

All five are optional. Each callback can return a Promise; the SDK awaits it before firing the next event.

Common uses#

• Forward metrics: in onLog, push step and loss to your own pipeline (PostHog, Datadog, etc.).
• Evaluate during training: in onCheckpoint, call infer() against a held-out prompt and log the sample so you know early if the run is heading in the right direction.
• Notify on completion: in onCompleted or onFailed, post to Slack or send an email.

For richer recipes, see Mid-run eval, Early stopping, and Notifications.

Reference#

For full type signatures, the rule that throwing inside a callback rejects wait() immediately (it is not routed to the SSE reconnect loop, which handles only transport failures), and the per-callback parameter list, see the Callbacks reference. For the conceptual flow of a run, see Run lifecycle.