The model must infer the expected behavior from the instruction and its pretrained knowledge. Prompt wording and output parsing remain part of the evaluation setup even though no demonstrations are supplied.
Zero-shot evaluation measures performance from task instructions without worked examples in the prompt.
The model must infer the expected behavior from the instruction and its pretrained knowledge. Prompt wording and output parsing remain part of the evaluation setup even though no demonstrations are supplied.