Skip to main content
Version: v0.23.x

AutoML

The AutoML section enables iterative fine-tuning using intelligent agent-driven configuration. You specify how many experiments you’d like to run, and the system automatically adjusts settings like learning rate and backbone selection to find the best-performing model.

If you’d rather get a single recommendation for one experiment at a time instead of running a full AutoML series, see Ask KGM. For an overview of both, see AutoML and Ask KGM in Key concepts.

AutoML runs table​

When you open the AutoML section from the top navigation bar or homepage, you’ll see a table of all AutoML runs within your project.

Each row shows:

  • Name: Searchable name of the AutoML run
  • Progress: Completed / total experiments
  • Best Score: Performance metric of the best experiment
  • Created: Date/time the run was started
  • Status: Completed, running, or queued
  • Actions: Rename or delete

Starting an AutoML run​

Click New AutoML to open the New AutoML Run page. The setup is similar to starting a regular experiment, and no manual hyperparameter configuration is required.

Under Backbone Selection, select the backbones for AutoML to evaluate. Select All selects the models under Available Backbones and Custom Backbones. If your administrator has configured a model hub:

  • Available Backbones lists the default models that are in the hub.
  • Other models in model hub lists all other models in the hub. Select All doesn't select these.

A notice above the lists explains why they might look different than expected, for example when the hub can't be reached.

Set the Number of Experiments (default: 5), then click Start AutoML. The system builds models iteratively:

  • Starts with a base config
  • Evaluates results
  • Adjusts parameters (for example, backbone or learning rate)
  • Starts the next experiment

Viewing an active run​

Once started, you’ll see:

  • Status: Queued, running, or completed
  • Progress: For example, 2/5
  • Best Score: The current best-performing configuration
  • Created: When the run was started
  • Stop / Delete: Controls for halting or removing the run

Experiments table​

This section displays:

  • Name of each individual experiment
  • Experiment ID
  • Dataset used
  • Created timestamp
  • Status: Completed, validating, or queued
  • Validation Metrics: For example, validation loss or perplexity

Click any experiment to open its detail page. See View and manage an experiment.

AutoML logs​

This expandable section shows a live, per-experiment progress feed:

  • Backbone models selected and parameter configurations tested
  • Status transitions (for example, an experiment moving from queued to running)
  • Periodic runtime, ETA, and train/validation loss updates while an experiment is in progress
  • A closing summary when an experiment completes (runtime and final validation loss), or its error message if it fails

Connected experiments​

When the run completes, a visual flow chart shows the iterative steps taken during AutoML, with each model, loss, and perplexity value for comparison.

Feedback

Related content

Experiments
Deployments