[Question] GPT zero/few-shot multiplechoice task evaluation implementation

[KTH1234]

Hello,

I noticed that in the NeMo GPT repository, the zero/few shot multiple-choice task evaluation scripts for RACE are evaluated using both "context + {choice candidate}" likelihood and "answer: {choice candidate}", as described in the GPT-3 paper. However, other similar tasks such as PiQA, BoolQ, and HellaSwag are only evaluated using "context + {choice candidate}".

I'm curious if there is a specific reason for this difference in evaluation method.