What is the meaning of score_gt and score_cand? about t-few HOT 6 OPEN

        if not self.config.split_option_at_inference:
            bs, num_choices = choices_ids.size()[:2]
            flat_choices_ids = choices_ids.flatten(0, 1)
            attention_mask = (input_ids != self.tokenizer.pad_token_id).float()  # [bs, max_seq_len]
            encoder_hidden_states = self.model.encoder(input_ids=input_ids, attention_mask=attention_mask)[0]
            encoder_hidden_states = encoder_hidden_states.unsqueeze(dim=1).repeat(1, num_choices, 1, 1).flatten(0, 1)
            attention_mask = attention_mask.unsqueeze(dim=1).repeat(1, num_choices, 1).flatten(0, 1)
            decoder_input_ids = torch.cat([torch.zeros_like(flat_choices_ids[:, :1]), flat_choices_ids[:, :-1]], dim=1)
            decoder_attention_mask = (decoder_input_ids == decoder_input_ids).float()
            lm_target = flat_choices_ids - 100 * (flat_choices_ids == self.tokenizer.pad_token_id).long()

            model_output = self.model(
                attention_mask=attention_mask,
                encoder_outputs=[encoder_hidden_states],
                decoder_input_ids=decoder_input_ids,
                decoder_attention_mask=decoder_attention_mask,
            )
            choices_scores = (
                F.cross_entropy(model_output.logits.flatten(0, 1), lm_target.flatten(0, 1), reduction="none")
                .view(bs, num_choices, -1)
                .sum(dim=-1)
            )
            if self.config.length_norm > 0:
                choices_scores = choices_scores / torch.pow(
                    (choices_ids != self.tokenizer.pad_token_id).sum(dim=-1), self.config.length_norm
                )
            pred_score, prediction = choices_scores.min(dim=1)

     score_gt = choices_scores[range(bs), labels]
     choices_scores[range(bs), labels] = choices_scores.max(dim=-1)[0]
     score_cand = choices_scores.min(dim=-1)[0]