Databricks Generative AI Engineer Associate Question 275
Select 3You have deployed a generative AI model for your organization, and users have started reporting that the model occasionally generates harmful or biased outputs. As part of your monitoring process, you need to systematically evaluate the model's behavior and implement safeguards. Which of the following actions should you prioritize to effectively monitor and address these issues?
- A
Implement automated tests to detect harmful or biased outputs on a curated evaluation dataset.
- B
Deploy the model in production without safeguards and rely on user feedback to monitor behavior.
- C
Set up a real-time monitoring pipeline to track the distribution of model outputs and flag anomalies.
- D
Periodically retrain the model without analyzing its current performance metrics.
- E
Use human-in-the-loop evaluations to review flagged outputs and refine the evaluation criteria.
Show answer and explanation
Correct answers: A, C, E
Explanation
Effective evaluation and monitoring of generative AI models require a combination of automated and human-driven approaches. Automated tests on evaluation datasets help catch issues early, while real-time monitoring ensures that anomalies are identified during production. Human-in-the-loop reviews provide additional safeguards by enabling nuanced evaluations and refining the monitoring process. Skipping safeguards or retraining without analysis can lead to unresolved or exacerbated issues.
- A. Correct.
Correct: Automated tests with a curated evaluation dataset are essential for identifying harmful or biased outputs before they become widespread in production.
- B. Incorrect.
Incorrect: Deploying the model without safeguards and relying solely on user feedback can lead to reputational damage and harm, as issues may not be detected promptly.
- C. Correct.
Correct: A real-time monitoring pipeline is critical for tracking model outputs and identifying anomalies or problematic patterns in production.
- D. Incorrect.
Incorrect: Retraining the model periodically without analyzing its current performance metrics may not address existing issues and can introduce new ones.
- E. Correct.
Correct: Human-in-the-loop evaluations allow for nuanced understanding of flagged outputs and help refine evaluation criteria for better monitoring.