AWS Certified Cloud Practitioner

Explanation

When a summarization LLM produces outputs that are too short and miss important details, the most direct solution is to increase the max_tokens parameter.

Why Option C is correct:

1. max_tokens controls output length: This parameter sets the maximum number of tokens the model can generate in its response. If outputs are too short, increasing this value allows the model to generate more detailed summaries.

2. Directly addresses the problem: The issue described is specifically about outputs being "too short and miss important details" - increasing token limit directly solves the length constraint.

Why other options are incorrect:

• Option A (Increase top-p): Top-p (nucleus sampling) controls diversity by limiting token selection to the smallest set whose cumulative probability exceeds p. Increasing it from 0.7 to 0.9 makes the model consider more diverse tokens but doesn't directly address output length.

• Option B (Reduce temperature): Temperature controls randomness - lower values (like 0.2) make outputs more deterministic and focused, but this might actually make summaries even more concise rather than more detailed.

• Option D (Increase repetition penalty): This parameter discourages repetition of words/phrases. While it might improve quality by reducing redundancy, it doesn't directly address the issue of outputs being too short.

Additional considerations:

• For summarization tasks, you might also consider adjusting the prompt to explicitly request more detailed summaries

• The min_tokens parameter (if available) could also be set to ensure a minimum output length

• Different models have different optimal token limits for summarization tasks