AWS Certified Cloud Practitioner

Explanation

Correct Answer: B - Decrease top-p to 0.6

Why this is correct:

1. Top-p (Nucleus Sampling) controls the cumulative probability threshold for token selection. When you decrease top-p (e.g., to 0.6), you restrict the model to consider only the most probable tokens that collectively make up 60% of the probability mass.

2. How it affects conciseness:

   • Lower top-p values make the model more deterministic and focused on high-probability tokens

   • This reduces the chance of the model wandering off into less relevant, verbose responses

   • The model becomes more conservative in its word choices, leading to more concise outputs

Why other options are incorrect:

• A. Increase temperature to 1.2: Higher temperature (above 1.0) increases randomness and creativity, which would likely make responses even more unpredictable and potentially more verbose.

• C. Increase top-k to 200: Higher top-k values allow the model to consider more token candidates (200 in this case), which increases diversity and could lead to longer, more varied responses.

• D. Disable repetition penalty: Repetition penalty helps prevent the model from repeating phrases. Disabling it might lead to repetitive or redundant content, not necessarily more concise responses.

Key Takeaway: For more concise outputs, you want to make the model more focused and deterministic, which is achieved by decreasing top-p to restrict the token selection to higher probability options.