Train or tune a model
Studio → Training runs real Vertex AI jobs. Each algorithm has a genuine prerequisite; jobs without it fail immediately with an actionable message instead of pretending to train.
Pick your algorithmlink
- check_circleFine-tune — supervised (adapter) tuning of a Gemini base model (the platform default base is currently gemini-3.1-flash-lite). Requires a gs://…/*.jsonl tuning dataset in the Dataset field.
- check_circleCustom — runs YOUR training container on Vertex (GPU selectable). Requires a container image URI; hyperparameters are passed as args (--dataset, --epochs, --learning-rate, --batch-size).
- check_circleAutoML — Vertex AutoML Tabular. Requires a Vertex-managed dataset id and a target column (both collected in the wizard).
Hyperparameter tuninglink
The tuning form runs a Bayesian/Random/Grid sweep over learning rate and batch size when you supply a training container — or, with just a gs:// JSONL dataset and base model, a single real supervised-tuning run. Trials and the best objective value sync back from Vertex as they complete.
Cost & statuslink
The wizard shows a deterministic Koin estimate before you submit. At dispatch, the server independently re-quotes the job from its parameters (compute type, epochs, algorithm) using the same rate table and escrows the server quote — so the price is server-authoritative; the wizard estimate normally matches it exactly. Escrow settles at completion and failed jobs refund in full. Status flows queued → running → completed/failed from live Vertex state.
Frequently asked questions
Why did my job fail instantly?expand_more
Almost always a missing prerequisite — the error on the job card states exactly what to provide (gs:// JSONL dataset, container image, or Vertex dataset id + target column).
Where does the trained model appear?expand_more
Fine-tuned/tuned models register in your Vertex project; the Studio Models page lists the models you register in Klick.