AIF-C01 Practice Question 6: Temperature Consistent Output
AWS Certified AI Practitioner (AIF-C01) · Test 1 · question 6 of 10
An insurance company uses a foundation model to pull the claim number, date, and amount out of claim emails. Testers report that the same email sometimes produces different wording or field order from one run to the next. Which inference parameter change is most likely to make the output more consistent?
Show answer and explanation
Answer:
At each step, a model picks the next token from a probability distribution. Temperature controls how that distribution is shaped. A lower temperature makes it steeper, so the model more reliably picks the most likely token and responses become more deterministic. That's what an extraction task needs (option D). Lowering top P or top K has a similar narrowing effect.
- A is wrong because a higher temperature flattens the distribution and makes lower-probability tokens more likely. That increases variation. - B is wrong because a higher top P widens the pool of candidate tokens. That also increases variation. - C is wrong because maximum length only caps how long the response can be. It doesn't affect run-to-run randomness.
Reference: Influence response generation with inference parameters