0–1 follows the model more/less strictly by frequency. Above 1 it starts swapping in words the model never saw here, and near 2 it can reach for words outside the model entirely.