Concept
Logits
Unnormalised scores a model emits over its vocabulary before softmax converts them to probabilities.
Definition
At every generation step a transformer produces a vector of logits â one per vocabulary token. A softmax turns them into a probability distribution from which the next token is sampled. Sampling parameters (temperature, top-k, top-p) reshape this distribution.
Common use cases
- Custom sampling
- Logit bias
- Constrained generation