Output Token Probabilities and Calibration
Warm-upDerivationprobabilitycalibrationlanguage models
Related module: Transformers and autoregressive models
Problem statement
Language models output probability distributions over tokens.
- For logits over tokens [A, B, C], compute the probabilities.
- What does it mean for a model to be calibrated?
- How can you estimate the probability that a model's answer is correct using its token probabilities?