引導式測驗 · 中等 等級
使用 Logprobs 評估 LLM 信心水準測驗
使用令牌對數機率作為特定於任務的訊號,並在將它們視為任何決策的答案置信度之前驗證校準。
問題 1 的 8
What does a token logprob describe?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What does a token logprob describe?
- The probability that an entire answer is factually correct
- The user’s confidence in a response
- The model’s conditional probability for a token given prior context
- The number of tokens in a prompt
- Why is a high token logprob not proof that an answer is correct?
- Logprobs always equal correctness labels
- A high value means the user approved the answer
- Tokens are not generated conditionally
- A model can assign likely tokens to a common but false continuation
- What affects a sequence score formed from token logprobs?
- Only the user’s screen size
- Tokenization, answer length, wording, and prompt design
- Only the model’s parameter count
- Nothing beyond the first token
- How should a team validate logprobs as a confidence feature?
- Define correctness, fit or choose a score on labeled data, then assess calibration on held-out examples
- Show the raw value to users without testing
- Use training examples as the only evaluation
- Assume scores are comparable across models
- What does calibration ask about a confidence estimate?
- Whether every token is spelled correctly
- Whether the model responds quickly
- Whether predictions assigned a confidence level are correct at the corresponding frequency
- Whether all outputs have equal length
- Why can a fixed-label classification task be easier to evaluate with logprobs?
- It makes probabilities automatically calibrated
- Candidate outputs are constrained and can be compared against labeled outcomes
- It removes all prompt sensitivity
- It guarantees no class imbalance
- What should be checked before using logprobs from an API?
- Only the model’s display name
- Whether the prompt asks for confidence
- Whether the network is wireless
- Whether the chosen endpoint and model support the parameter and returned values
- How should a system handle a low-confidence high-impact case?
- Route them to review or a validated fallback
- Accept every score as truth
- Hide the score and proceed automatically
- Increase output length without evaluation