Questions about training data preprocessing

[liuqi6777]

Hi!

I noticed that "attention_mask" was ignored when preprocessing the training data, as shown in the code of the file src/tevatron/data.py.

class TrainDataset(Dataset):
    def create_one_example(self, text_encoding: List[int], is_query=False):
        item = self.tok.prepare_for_model(
            text_encoding,
            truncation='only_first',
            max_length=self.data_args.q_max_len if is_query else self.data_args.p_max_len,
            padding=False,
            return_attention_mask=False,
            return_token_type_ids=False,
        )
        return item

And I found that some other sources of work on dense retrieval didn't do this. So I want to ask what is the reason for designing the code like this.

Thanks for your answer:)

[liuqi6777]

I just found that the QPCollator will add "attention_mask" to it.