Logit Bias Quiz

Answers and explanations appear as you take the quiz.

1. What does logit bias modify?
   • The score (logit) of specific tokens before sampling
   • The model's training weights permanently
   • The temperature of the whole prompt
   • The number of tokens in the context window
2. In many APIs, what does a logit bias of -100 on a token accomplish?
   • Makes the token 100 times more likely
   • Effectively forbids the token from being generated
   • Deletes the token from the vocabulary forever
   • Doubles the model's speed
3. What numeric range does the documented OpenAI Chat Completions logit_bias accept for a token?
   • From 0 to 1 only
   • From -1 to 1 only
   • From -100 to 100
   • Any string value
4. Logit bias is keyed to what identifier?
   • The plain-text word
   • The sentence position
   • The model's name
   • The numeric token ID
5. Because logits are passed through a softmax, what is the effect of adding a positive bias?
   • It exponentially boosts the token's relative weight before normalization
   • It linearly adds a fixed probability
   • It has no effect until training
   • It removes the token from consideration
6. A key advantage of logit bias over fine-tuning is that it:
   • Permanently improves the model's knowledge
   • Requires no retraining and applies per request
   • Increases the context window size
   • Works only on images
7. Why should a developer inspect the tokenizer before applying a bias to a visible word?
   • Token IDs are assigned only after sampling
   • Visible words always equal one token
   • The word may be represented by several token IDs, so biasing one ID may not cover all forms
   • The tokenizer controls the model’s training data
8. What is a practical risk of applying a strong positive bias to one token?
   • It increases the context window
   • It changes the model weights permanently
   • It guarantees the token appears exactly once
   • It can distort generation or produce repetitive, awkward text