A unidirectional language model predicts tokens using only the tokens that come before the target position. This contrasts with bidirectional models, which can use both preceding and following context.
A language model that bases its probabilities only on the tokens appearing before , not after , the target token(s). Contrast with bidirectional language model.