Logits
Also known as: logit
Logits are the raw output numbers a model produces for each token in the vocabulary — before they’re converted into probabilities.
The pipeline:
- Model computes logits (raw scores).
- Temperature is applied to the logits.
- Softmax turns them into a probability distribution.
- Sampling picks a token.
Why you'd care: "logit bias" is a feature in many APIs that lets you manually boost or suppress specific tokens. It's a blunt but sometimes very useful tool.