Setting the no_answer probability in the squad_v2 metric

Hi, I just wanted to clarify whether the no_answer_probability that we need to pass to the squad_2 metric (in the predictions dictionary) should be equal to only the sum of start and end logits for the CLS token (i.e start_logit[0]+end_logit[0]) or should it be the difference start_logit[0]+end_logit[0]-best_start_logit-best_end_logit. Here start/end_logit[0] refer to the scores of the CLS token.