Question 320
At fictional Redfern Regional Insurance, a Snowflake data and AI engineering team is conducting a multi-tenant knowledge assistant improvement using only first-party public technical references and independently written facts. The separately composed technical incident is: A language team fine-tunes a model on customer interaction samples, but it lacks a held-out evaluation and cannot tell whether domain gains offset losses in safety and general performance. Assume only these documented facts and use safe, authorized production-change procedures; exact current Snowflake feature syntax and region or account availability require current review. Which repeatable assurance or control is most relevant to preventing this situation?
Explanation
Writer-proposed answer: D (UNVERIFIED)
Preliminary individual option explanations — NOT expert validated
Correct Answer
Document data approvals, deduplication rules, holdout performance and a rollback route for every fine-tuned model release. Fine-tuning is an optimization choice requiring independent evaluation and data governance, not automatically better performance.
Incorrect Answers
This response concerns Sentiment interpretation with irony, which is different from the stated failure. A language team fine-tunes a model on customer interaction samples, but it lacks a held-out evaluation and cannot tell whether domain gains offset losses in safety and general performance.
Similarity threshold excludes rare queries is insufficient to demonstrate this result: A controlled evaluation shows improvements on relevant unseen cases without unacceptable safety regressions, with the trained artifact and dataset lineage recorded.
Sensitive translation with regional constraints is not the cause of the reported failure: A language team fine-tunes a model on customer interaction samples, but it lacks a held-out evaluation and cannot tell whether domain gains offset losses in safety and general performance.