why the target_label is 0 for all training dataset?

[hecongqing]

class RerankerModel(nn.Module):
    TRANSFORMER_CLS = AutoModelForSequenceClassification

    def __init__(self, hf_model: PreTrainedModel, train_batch_size: int = None):
        super().__init__()
        self.config = hf_model.config
        self.hf_model = hf_model
        self.train_batch_size = train_batch_size
        self.cross_entropy = nn.CrossEntropyLoss(reduction='mean')
        if train_batch_size:
            self.register_buffer(
                'target_label',
                torch.zeros(self.train_batch_size, dtype=torch.long, device=self.hf_model.device)
            )
        for name, param in self.hf_model.named_parameters():
            # for some reason, ds zero 3 left some weights empty
            if 'modules_to_save' in name and param.numel() == 0:
                logger.warning(f'parameter {name}, shape {param.shape} is empty')
                param.data = nn.Linear(self.hf_model.config.hidden_size, 1).weight.data
                logger.warning('{} data: {}'.format(name, param.data.cpu().numpy()))

[MXueguang]

sorry for the late reply,

we do this because we place positive doc at index 0, the following docs are negatives