Explanation
Correct Answer: B (Increase top-p)
Why this is correct:
- Top-p (nucleus sampling) controls the diversity of vocabulary by selecting from the smallest set of tokens whose cumulative probability exceeds the threshold p.
- Increasing top-p (e.g., from 0.7 to 0.9) allows the model to consider a broader range of possible tokens while still maintaining high probability tokens, which helps achieve a slightly broader vocabulary without losing meaning.
- This approach is more controlled than temperature adjustments because it maintains probability-based selection rather than just flattening the distribution.
Why other options are incorrect:
- A. Decrease temperature: Lowering temperature makes the model more deterministic and less creative, which would make the text even more rigid (opposite of what's needed).
- C. Reduce top-k: Reducing top-k limits the number of tokens considered at each step, which would make the vocabulary even more constrained.
- D. Increase max-tokens: This only affects the maximum length of generated text, not the vocabulary diversity or creativity.
Key Concepts:
- Temperature: Controls randomness (higher = more random, lower = more deterministic)
- Top-p (nucleus sampling): Controls vocabulary diversity by probability threshold
- Top-k: Limits the number of tokens considered at each step
- Max-tokens: Controls maximum output length
For the specific requirement of "slightly broader vocabulary without losing meaning," increasing top-p is the most appropriate adjustment as it allows for more lexical variety while maintaining semantic coherence.
