Limited Availability
← Gnok home

Train models and compare predictions

Train a model on your tables, import an ONNX model, or connect one you already host. Track which version scored each invoice and compare its estimate with how many days the customer took to pay.

For identical inputs, the initial ONNX model scores 0.694873 and the activated candidate scores 0.789680.
Recorded scores from two versions of the same synthetic fraud model. Both used the same four input values.

Train from a table

The churn example uses tenure, monthly charges, support tickets and contract length. A logistic regression trained on eight synthetic customers returns a score of 0.9413 for the example customer. Saving the feature definitions with the model lets the scoring query apply the same units and scaling.

Train and score the churn model

Compare two model versions

Register an ONNX model and declare its inputs and return type. Stage a new version under the same name, then compare the scores both versions give the same input.

For a days-to-pay model, save the predicted days, invoice ID and model version together. When payment dates arrive, compare actual and predicted days for matching invoices. Mean absolute error tells the finance team how far off the estimates were, in days.

Try a staged ONNX rollout

Call an existing endpoint

If you already host a model, register its endpoint, inputs and output in Gnok. The tutorial sends three numbers to the scoring service and checks that it returns the expected score of about 0.729088.

Test a remote model