Katib scenarios

The following hypothetical examples show how hyperparameter tuning can improve model performance:

• An image classifier reaches 82% validation accuracy with its initial configuration. Katib tests learning rates, batch sizes, and optimizers; the best trial reaches 91%.

• A training job starts with a validation loss of 0.42. After Katib tests combinations of learning rate and momentum, the best trial reduces the loss to 0.27.

Katib automates the repetitive work of testing hyperparameter combinations and comparing their results. Use these scenarios to validate your Katib setup before launching long-running experiments. They cover:

• successful experiment execution,

• hyperparameter tuning, metric collection, and best-trial selection,

• parallel trials with resource limits,

• behavior under resource pressure,

• notebook-based execution,

• namespace isolation, and

• UI-based experiment creation.

In this way, you will be able to verify that trials start, respect resource limits and namespace boundaries, and report results correctly in MLOps Studio.

You can run them in several ways:

• from a local terminal by using the Katib Python client,

• by applying experiment manifests with kubectl,

• from a notebook running in MLOps Studio,

• directly from the MLOps Studio UI.

Each scenario is a validation procedure with a goal, required setup, actions, and expected result.

• Terminal and kubectl Katib validation
• Notebook-based Katib validation
• Katib UI and namespace isolation validation
• Katib UI hyperparameter tuning

What to do next

Continue with the guide that matches how you plan to run or automate later experiments:

• Download and use kubeconfig for a MLOps Studio namespace

• Run Python code in a MLOps Studio notebook environment

• Use notebooks with MLOps Studio Pipelines