`createTrainer`

createTrainer is where you describe a fine-tuning run: which base model, which dataset, what knobs. The result is a Trainer that arkor start (and Studio’s Run training button) drives.

import { createTrainer } from "arkor";

export const trainer = createTrainer({
  name: "support-bot-v1",
  model: "unsloth/gemma-4-E4B-it",
  dataset: { type: "huggingface", name: "arkorlab/triage-demo" },
  lora: { r: 16, alpha: 16 },
  maxSteps: 100,
});

The fields you reach for first

name: shows up in Studio and in cloud-side logs. Pick something specific.
model: the base open-weight model. Templates use gemma-4-E4B-it.
dataset: where the training data lives. See DatasetSource.
lora: LoRA / QLoRA knobs. r: 16, alpha: 16 is a fine default; omit to take the backend default.
maxSteps or numTrainEpochs: cap how long the run goes.
callbacks: see Callbacks.

Try it without a real run

dryRun: true tells the backend to truncate the dataset and cap steps so a run finishes in a couple of minutes while still exercising every stage of the pipeline. Useful when wiring up callbacks for the first time.

createTrainer({
  name: "smoke",
  model: "unsloth/gemma-4-E4B-it",
  dataset: { type: "huggingface", name: "arkorlab/triage-demo" },
  dryRun: true,
});

Reference

For the full TrainerInput shape, every typed optional field, LoraConfig, the unstable forwarded fields (warmupSteps, loggingSteps, saveSteps, evalSteps, etc.), and the multi-trainer roadmap note, see the createTrainer reference.