Sampling

Choosing the next token from the model's probability distribution rather than always taking the top one. Settings such as temperature, top-k and top-p shape it.

All terms, A to Z