Skip to main content
In a few minutes you will go from zero to a fine-tuned model you can chat with in a local Playground. The training run itself takes 7 to 12 minutes. The first job you submit also waits for a GPU to be allocated before the training run starts; that wait can be a few minutes (occasionally longer) and only happens for the first run after a quiet period.
How Arkor uses GPUs. Arkor splits GPUs into two pools:
  • Inference GPUs run the model after training (the Playground and any production inference). These stay on an active worker, so they are always warm and respond immediately.
  • Training GPUs run fine-tuning jobs. To keep costs predictable, these are not kept warm. When you click Run training, Arkor allocates a GPU on demand. If a worker is still warm from a recent job, the allocation finishes in seconds. Otherwise it starts from cold (the so-called cold start: provisioning, pulling the container image, loading the base model), which can take several minutes.
After GPU allocation, the training run itself takes about 7 to 12 minutes for the templates. The Jobs view in Studio shows both phases.

​
Prerequisites

  • Node.js 22.22.0 or newer. Node 24 is recommended.
  • A package manager. pnpm, npm, yarn, and bun all work. Each snippet below ships as a tab group, so pick the one you use and the rest of the page follows.
  • A working internet connection. Training runs on Arkor’s managed GPUs.
No account, no credit card, no GPU on your machine.

​
1. Scaffold a project

pnpm create arkor my-arkor-app
cd my-arkor-app
The scaffolder asks which template you want. Pick the one closest to what you eventually want to build:
TemplateTaskOutput shapeTraining run (excludes GPU allocation)
triageSupport triage{ category, urgency, summary, nextAction }~7 min
translate9-language translation{ translation, detectedLanguage }~7 min
redactionPII redaction{ redactedText, redactedCount, tags }~12 min
Each template pairs the same small open-weight base (gemma-4-E4B-it) with a curated public dataset on HuggingFace. The training is real and finishes in minutes, so you get to see the whole loop end to end. To skip the prompt: 
pnpm create arkor my-arkor-app --template triage

​
2. Look at what was generated

my-arkor-app/
├── src/arkor/
│   ├── index.ts        # createArkor({ trainer })
│   └── trainer.ts      # createTrainer({ ... })
├── arkor.config.ts
├── .arkor/             # state + build artifacts (gitignored)
└── package.json        # dev / build / start
The two files that matter for now:
  • src/arkor/trainer.ts holds your trainer definition. Model, dataset, LoRA settings, hyperparameters, and lifecycle callbacks all live here.
  • src/arkor/index.ts is the entry point. It calls createArkor({ trainer }) so the CLI and Studio can discover what to run.
Open trainer.ts. The shape is the same across templates; only the dataset, output schema, and example prompts differ.

​
3. Open Studio and start a run

pnpm dev
This runs arkor dev, which boots Studio at http://localhost:4000. Studio is the UI; arkor dev itself does not start training and does not watch your trainer files. In the browser, click Run training. Studio submits the job to the managed backend in the background and streams the output back into the page. The first time you trigger anything, an anonymous workspace is created automatically: no signup, no credit card. Once a run is in flight, three views matter:
  • Jobs. A list of training runs. Click into one to see live status.
  • Loss chart and event log. As the run streams from the managed GPU, the loss curve updates and the log tail shows training events. The first job’s first phase is the GPU allocation; during that time the status reads Warming up GPU, the loss chart shows a Waiting for GPU placeholder, and the events list is empty. Once the training.started event arrives, the status flips to Running and the curve begins to update. The training run itself takes about 7 to 12 minutes depending on the template.
  • Playground. After a job completes, pick the final adapter from the selector and chat with it. Use the mode toggle to switch between the base model and the adapter. To run inference on intermediate checkpoints while a run is still in flight, use onCheckpoint callbacks instead of Studio.
If you edit src/arkor/ between runs, refresh the Run training page (or run arkor build) before the next click so the new code is what runs.

​
4. Tie runs to an account (optional)

Anonymous workspaces are useful for trying things out, but the work is tied to the anonymous token on this local machine and is not visible from another device or another account. To keep training under an account, sign in with OAuth before you start the next run: 
pnpm arkor login --oauth
This runs the Arkor Cloud OAuth (PKCE) flow on a loopback port and links your local credentials to an account. From then on, runs from this machine show up under that account on any device. The scaffolded project installs arkor as a local devDependency, so use your package manager’s local-bin runner over a bare arkor invocation unless you have installed the CLI globally. The generated project README walks through the equivalent setup for pnpm, npm, yarn, and bun. A few caveats. Running arkor login without a flag opens an interactive picker that defaults to Anonymous; accepting the default mints a brand-new anonymous token (a new anonymousId) and overwrites ~/.arkor/credentials.json, so it lands you in a different workspace from the previous anonymous session, not a refresh of it. Switching to OAuth overwrites the same file and does not migrate prior anonymous workspaces or jobs into the account. Merging anonymous work into an OAuth account once you sign in is on the roadmap; until that lands, sign in before the runs you want associated with the account.

​
5. Where to go next

  • Concepts. Read Concepts to build a mental model of createArkor, createTrainer, the lifecycle callbacks, and Studio.
  • Customize the trainer. Open src/arkor/trainer.ts and tweak lora.r, maxSteps, or add more callbacks. Refresh the Run training page (or run arkor build) before the next click so your edits land.
  • Try another template. Re-run the scaffolder with a different --template to compare.