SnowPro Specialty: Gen AI Question 51
Single answerFine-tuningA retail company uses a general-purpose large language model in Snowflake to generate customer support replies. The model is fluent, but it often misses the company's approved refund wording and occasionally uses phrases that violate internal compliance guidelines. The team has a curated dataset of past approved responses reviewed by legal and support leads. They want the model to produce outputs that better match this company-specific style and policy language for this support use case. Which approach is the MOST appropriate?
- A
Fine-tune the model using the curated, high-quality approved support response dataset so the model better learns the company-specific patterns and wording
- B
Increase the temperature during inference so the model explores more possible phrasings and is more likely to match the approved refund language
- C
Replace the curated approved-response dataset with a much larger unlabeled corpus of all support chats, because volume matters more than label quality for this requirement
- D
Avoid any model adaptation and rely only on longer prompts that restate the policy, because fine-tuning is intended only to improve factual knowledge retrieval
Show answer and explanation
Correct answer: A
Explanation
The best answer is to fine-tune the model on a curated set of approved support responses. This scenario is not primarily about adding external knowledge; it is about making outputs consistently reflect organization-specific language, compliance constraints, and response patterns. Fine-tuning is well suited for adapting model behavior to specialized tasks when high-quality example pairs or reviewed outputs are available. By contrast, inference parameter changes such as temperature only alter sampling behavior, not learned policy wording. Likewise, replacing curated data with a larger but noisier corpus is risky because low-quality or inconsistent examples can degrade alignment. Prompting remains useful, but it is not the most appropriate standalone solution when the requirement is persistent adaptation to a narrow communication style. This aligns with common generative AI best practices: use fine-tuning for behavior/style adaptation and use retrieval when the main problem is grounding responses in current facts or enterprise data.
- A. Correct.
Correct. Fine-tuning is the most appropriate approach when the goal is to adapt a base model's behavior, style, phrasing, and task performance to a domain-specific use case using curated examples. In this scenario, the company has a reviewed dataset of approved responses, which is exactly the kind of supervised data that can help a model better align to required refund wording and compliance-sensitive phrasing. This is a practical use case for fine-tuning: improving consistency of outputs for a specific business workflow.
- B. Incorrect.
Incorrect. Increasing temperature affects randomness and creativity at inference time; it does not teach the model company-specific policy wording. In fact, higher temperature generally makes outputs less deterministic and could increase variation, which is undesirable in a compliance-sensitive support scenario. A candidate might choose this option because generation settings do influence tone, but they are not a substitute for adapting the model to approved examples.
- C. Incorrect.
Incorrect. More data is not automatically better if it is unlabeled, inconsistent, or contains non-approved responses. For this scenario, quality and alignment of the training examples matter more than simply increasing corpus size. Using all support chats could introduce undesirable language patterns, outdated policy wording, or noncompliant responses, which would undermine the goal of aligning model outputs to approved standards.
- D. Incorrect.
Incorrect. Prompting can help guide responses, but the statement that fine-tuning is only for improving factual knowledge retrieval is a misconception. Fine-tuning is commonly used to adapt behavior, style, formatting, and domain-specific task performance. Retrieval-based approaches are more appropriate when the issue is accessing up-to-date facts or external knowledge, whereas this scenario is primarily about response style, approved wording, and policy-aligned behavior.