IBM Watson OpenScale requires access to the model's training data to establish a baseline for its monitors. This training data is used to generate essential statistics that form the foundation for detecting model drift and evaluating fairness. For instance, OpenScale analyzes the training data to create a drift detection model, which compares the statistical properties of runtime data against the training data's characteristics. Similarly, for fairness monitoring, OpenScale uses the training data to understand the distribution of protected attributes and outcomes to identify potential biases in the model's predictions.

A. Test data subset (10%): A small subset is insufficient for creating the robust statistical baseline required for drift detection and fairness analysis.

B. Complete test data set: While test data is crucial for pre-deployment model validation, OpenScale specifically uses the training data as the reference to monitor the live, deployed model.

D. Access to Pipeline: Access to a CI/CD pipeline is related to MLOps automation (building and deploying models), not the post-deployment monitoring and statistical analysis performed by OpenScale.

1. IBM Cloud Docs

"Configuring models for monitoring in Watson OpenScale": In the "Provide model details" section of the configuration wizard

it explicitly states the requirement to connect to the training data. "To configure monitors

Watson OpenScale requires access to your training data which you can store in a Db2 database or Cloud Object Storage." This confirms that training data is a prerequisite for configuration.

2. IBM Cloud Docs

"Configuring the drift monitor": This document details the process for setting up drift detection and states

"To detect drift

Watson OpenScale builds a drift detection model... The drift detection model processes your training data to create a baseline." This directly links training data to the generation of drift statistics.

3. IBM Cloud Docs

"Preparing your training data": This page outlines the requirements for the training data that will be provided to OpenScale. It notes

"For Watson OpenScale to analyze your model

you must provide access to the training data that you use to train your model." This reinforces that training data is the fundamental input for analysis.