Mental model

The mental model behind Arkor: project layout, trainer, lifecycle callbacks, and Studio.

This section gives you the mental model you need before diving into the SDK reference or CLI reference. Read these in order; each builds on the previous.

What to read#

• Project structure. Where Arkor expects code to live (src/arkor/index.ts, src/arkor/trainer.ts), what .arkor/ and ~/.arkor/ are for, and what arkor.config.ts does (and does not) do today.
• Trainer. The shape of createTrainer: required fields, dataset sources, LoRA settings, and the Trainer.start / wait / cancel lifecycle.
• Lifecycle callbacks. The five callbacks (onStarted, onLog, onCheckpoint, onCompleted, onFailed), their argument shapes, and when each fires. This is where most of the value of doing fine-tuning in TypeScript lives.
• Studio. What the local web UI is for, why it runs on loopback, and how it fits into the dev loop.

The one-paragraph version#

You define a Trainer with createTrainer({ name, model, dataset, ... }) in src/arkor/trainer.ts and register it through createArkor({ trainer }) in src/arkor/index.ts. arkor dev opens a local Studio in the browser; clicking Run training in Studio spawns arkor start, which calls trainer.start() and trainer.wait(). While wait() is open, callbacks dispatch from the SSE event stream, training streams from Arkor's managed GPUs, and the same view is reflected back into Studio.