Question 210
A batch command job performs differently each run despite identical data due to randomized weight initialization. What should be recorded in code and metadata?
Explanation
Correct Answer
Controlled randomness reduces accidental variation and identifies which seeds were used for each model artifact.
Incorrect Answers
Log or publish the evaluation metric through the documented run output interface. This is useful for the distinct issue: A training job completes but a release gate cannot read the evaluation score from printed output. The relevant technical rationale or evidence for that issue is: Structured metrics and artifacts allow release automation to consume evidence without fragile parsing of console text. Here the reported issue is: A batch command job performs differently each run despite identical data due to randomized weight initialization. The relevant response is: Set and log reproducible random seeds for the relevant training libraries. The first result cannot verify this different correction: Controlled randomness reduces accidental variation and identifies which seeds were used for each model artifact.
Use a pinned versioned Azure ML training environment instead of ad hoc installs. It addresses another case: A training script installs random package versions when it starts and later produces irreproducible runs. For that case, the distinguishing justification or acceptance check is: A managed environment describes fixed dependencies before execution and supports repeatable training containers. The present case instead states: A batch command job performs differently each run despite identical data due to randomized weight initialization. It calls for Set and log reproducible random seeds for the relevant training libraries. Showing success for the other case would not establish: Controlled randomness reduces accidental variation and identifies which seeds were used for each model artifact.
The alternative, Use the Azure ML CLI v2 ml extension and compatible job YAML syntax, belongs to a different technical condition: A CLI automation uses old azureml experiment commands and cannot submit a current v2 command job definition. There the source-specific mechanism or proof is: Current Azure ML job submission uses v2 resource commands and definitions rather than obsolete v1 interfaces. The condition asked about here is: A batch command job performs differently each run despite identical data due to randomized weight initialization. A suitable correction is Set and log reproducible random seeds for the relevant training libraries; the targeted justification or result is: Controlled randomness reduces accidental variation and identifies which seeds were used for each model artifact.