Skip to main content

The Model Page

Open a model from the Model Registry to see its page: the model's card, with everything about it in one place. The header shows its name (with an AutoML badge for a published AutoML model) and, for those who may change it, Deploy, Start, Stop and Delete. Below it are the model's Status, Framework and when it was Created. Its tabs follow the order every detail page uses: Overview, Permissions, the model's own tabs, its configuration, what it has done, and Logs last.

The open tab is kept in the page's address (?tab=monitoring, for example), so a reload or a shared link opens the same tab.

Overview​

Model Details: name, type, framework, algorithm, Task, evaluation metric, training data, size, tags, and when it was created and updated. A model without a task shows Not set: drift cannot be measured until its owner sets it (see Monitoring).

Feature Information: the target variable and the model's input features.

Permissions​

Who the model is shared with: named users, or all users, each to use it or to edit it. See Access.

Versions​

Every version, with its status (Active & Running, Active or Inactive), description, size and when it was created, and Add Version, Set as Active, Deploy this version and Delete version. See Versions.

Features​

The model's input features, each with its Type, whether it is Required, and its Importance when the model has feature importance.

Feature Service Binding connects the model to a feature store so callers send only entity keys and the model resolves its features at inference:

  • A model trained on a feature service shows its binding (Trained on the store and service version) and its Feature lineage: each feature and the view it comes from.
  • Any other model is bound here: choose the Feature Store and Feature Service, turn on Resolve online features at inference, optionally set Max staleness (seconds), and select Save binding. Remove binding undoes it.

See Models and features.

Test​

Available while the model is running, or idle on demand (its first prediction wakes it). Enter a value for each input, typed by the model's input schema, and select Predict. The result shows the Prediction, Class Probabilities for a classifier, the model's whole Output, and which version Served by. For a multimodal model's image input, pick an image file.

Test predictions are recorded and marked Test: they never count in drift or performance results.

Environment​

How the model is served: its resources, replicas, scheduling mode and autoscaling. See Deploy and scale.

Metrics​

The model's offline evaluation metrics (for a classifier, its confusion matrix and per-class precision, recall and F1 when it has them) and its Training Info: the experiment and run, or the AutoML job, it came from, the hardware it trained on, when training started and completed, the time limit, and its hyperparameters.

Monitoring​

The model's monitoring settings, drift status, predictions, actuals imports, performance and baselines, with Build Baseline, Import Actuals, Run Analysis and Full Dashboard. See Monitoring.

Analytics​

Who calls the model, how often, and how fast and well it answers, over 7 days, 30 days or 90 days:

  • Model usage: the period's Predictions, Failed predictions (a tiny share reads as <0.1%), Latency and Callers.
  • Predictions per day and Latency per day.
  • Versions: per model version (Unversioned for a model served outside the platform), its predictions, failures, median latency, confidence and when it was last called.
  • A/B test variants: predictions and failures per A/B test and variant.
  • When it is called (UTC): a heatmap of calls by weekday and hour.
  • Callers: each caller (or Anonymous), Through what they called (Direct (API or UI), an App, a Workflow or a Workspace, with its name), their predictions, failures and when they Last called.

Logs​

The model's build logs and its runtime (container) logs. A failed build or start says why here.