Skip to main content

Models in Studio

Open Operations → Models. Each row represents a registered version. Check the version and status before inspecting or scoring it. Studio uses your selected connection and workspace permissions.

Open a model version​

  • Model name opens a read-only page for that exact registered version. Bookmark the page or share its URL with a teammate who has access to the workspace and Studio.
  • Chevron expands the row to show version actions, lineage, and evaluation history. Open the Workflow tab to inspect the same graph inline.
  • Workflow on the row opens the graph directly. Open model page takes you to the dedicated page.

The workflow connects the recorded data, features, training and evaluation results, and scoring and outcome jobs for the version. Table nodes can show a fully qualified name and a pinned snapshot. Select a table node, then Open table in Catalog, to inspect it. Feature nodes show model input order and the saved SQL. Choose a run node to see its status, SQL, and results. A job card shows its latest run. Use View run history for earlier runs.

Use Expand canvas for more room, the zoom buttons to change scale, and drag the canvas background to pan. Fit to view brings the full graph into the viewport. The dedicated page opens expanded and fitted. Refresh the workflow to load the latest recorded status.

Import a model or train one​

Start hereUse it to
Upload artifactRegister a compatible .onnx file. Match the declared argument types, order, and return type to the artifact. Use a qualified model name and verify representative predictions after registration.
Train new → AutoMLTrain from a SQL query with a built-in estimator or saved training type.
Train new → PythonTrain from a SQL query using a Python script in the Studio training sandbox and export its model to ONNX.
ExperimentsCompare estimators on a shared evaluation split, keep training evidence, and select a trial for registration.
New modelOpen the SQL editor with a native CREATE MODEL template.

For an artifact's required tensor shapes and SQL signature, see ONNX inference.

AutoML's built-in training choices include linear regression, Ridge, Elastic Net, logistic regression, random forests, and gradient boosting. The first training run can take longer while the training worker starts.

Train a sample invoice model​

Use a model name in a schema where you can register models, such as tutorial.dashboard_samples.invoice_days_demo. With Train new → AutoML, enter the name and the following synthetic Training Data (SQL Query).

SELECT
CAST(100 + (n % 9) * 100 AS REAL) AS invoice_amount_usd,
CAST(n % 5 AS REAL) AS prior_late_payments,
CAST(15 + (n % 9) * 2 + (n % 5) * 3 AS REAL) AS actual_days_to_pay
FROM generate_series(1, 80) AS invoices(n);

Choose Preview Columns. Select actual_days_to_pay as the target and verify that the features are invoice_amount_usd and prior_late_payments, in that order. The wizard excludes the selected target from the feature list. Choose Ridge or Elastic Net, then Train Model. This screen trains and registers the result.

In the SQL editor, replace the function name if you chose a different schema or model name. This query applies the model to three sample invoices.

SELECT n AS invoice_id,
tutorial.dashboard_samples.invoice_days_demo(
CAST(250 * n AS REAL),
CAST(n - 1 AS REAL)
) AS predicted_days_to_pay
FROM generate_series(1, 3) AS invoices(n);

An active model with training settings saved by Studio has an Edit training action. It opens the saved SQL and training configuration. For a candidate comparison and review before activation, use Experiments.

Save a reusable Python training type​

In Train new → Python, enter the training SQL and edit the Python script. The starter script reads the SQL rows as the pandas DataFrame df and assigns the exported artifact to onnx_model. Use a numeric feature order and an output shape that match your ONNX inference signature.

Choose Save as AutoML type and name the definition. The saved SQL column order defines the ordered features, with the target in the final column. In AutoML, select the type from Saved Python types and choose Use saved SQL to load its query. Preserve the saved column names and order when supplying a replacement query.

On Models, use Saved AutoML types to View, Edit, or Delete a type. Deleted types lists recoverable definitions. Edits and restores apply to the reusable definition used in future training runs.

Compare and reproduce training​

  1. Open Experiments → New experiment. Supply a qualified model name and a read-only training query, then Preview columns.
  2. Choose the outcome and ordered inputs. Set the holdout, seed, and ranking metric. Use a time split when the model should train on older events and be tested on later ones. For tuning, choose validation folds and keep the final test set separate.
  3. Choose one or more estimators and Start experiment. Compare their holdout or validation metrics and the baseline. The result can also show calibration and importance when you enable the applicable options.
  4. Inspect a trial's View code, data profile, and Run settings and provenance before registering it.

Pin sources and features​

In Reproducible inputs, enable Pin training table snapshots. Enter a qualified table, choose a SQL alias, and use Find snapshots and Add pinned source. Reference the alias in your training SQL. Return a unique, non-null ordering column and set it as Unique order column so Studio can reproduce the row order.

For point-in-time features, select Look up feature group. Inspect the event columns and map the event ID, event timestamp, and entity keys. Supply the history timestamp column and select the features. Use Find feature history snapshots to pin the history query. Use point-in-time SQL inserts the query that selects the latest matching feature value at or before the event timestamp.

A completed pinned experiment shows the inputs and their SHA-256 digest in Run settings and provenance. Choose Rerun from these inputs to reuse the saved estimator script and worker image and verify the input digest.

Register and review a version​

From a completed experiment, choose Stage candidate for an existing model. The current active version continues serving while you review the candidate. For a new model, Studio offers Register as first active version and explains that registration will make the first version active.

Expand the model row and choose Release review on the candidate. Supply a representative Replay SQL query with the model features and actual outcomes, then choose the task and set the maximum allowed regression. Review the overall and segment results. When Studio's review policy requires approval, a second workspace administrator reviews the passing replay before promotion. Use the version's explicit Promote or Roll back action to choose the active version.

For scoring previews, version contracts, batch jobs, and delayed outcomes, see Model operations.