AWS Certified Cloud Practitioner

Explanation

For a financial compliance chatbot that requires factual and deterministic answers, the key parameters to consider are:

Temperature Parameter

• Temperature controls the randomness/creativity of the model's output

• Lower temperature (closer to 0) = more deterministic, focused, and factual responses

• Higher temperature (closer to 1 or above) = more creative, diverse, and random responses

Top-p (Nucleus Sampling)

• Top-p controls the cumulative probability threshold for token selection

• Lower top-p values = more focused on high-probability tokens

• Higher top-p values = more diverse token selection

Analysis of Options:

A. Temperature = 0.1, top-p = 0.5 ✓ CORRECT

• Temperature 0.1 is very low → highly deterministic

• Top-p 0.5 is moderate → focuses on high-probability tokens

• This combination produces consistent, factual responses

B. Temperature = 1.0, top-k = 100

• Temperature 1.0 is high → introduces randomness

• Top-k 100 allows many tokens → increases diversity

• Not suitable for factual compliance responses

C. Temperature = 0.9, top-p = 0.95

• Temperature 0.9 is high → too much randomness

• Top-p 0.95 is very high → includes many low-probability tokens

• Creates creative but potentially inaccurate responses

D. Temperature = 1.2, enable repetition penalty

• Temperature 1.2 is very high → maximum randomness

• Repetition penalty prevents repetition but doesn't ensure factual accuracy

• Worst choice for factual compliance chatbot

Why Option A is Best:

1. Low temperature (0.1) minimizes randomness and ensures the model chooses the most likely, factual tokens

2. Moderate top-p (0.5) focuses on high-probability tokens without being overly restrictive

3. This combination produces consistent, reliable, and deterministic responses essential for financial compliance

4. Financial compliance requires zero creativity and maximum factual accuracy

Key Takeaway:

For applications requiring factual, deterministic outputs (compliance, legal, medical, technical documentation), always use low temperature values (0.1-0.3) and moderate to low top-p values (0.5-0.7) to ensure consistency and accuracy.