Hi - from the paper it mentions proportions of 1,2,3-step questions in the dataset. However, there is no explicit label indicating which question is a 1-step question or a 2,3-step question. Could you add that information or provide a way to infer the number of hops a question takes given a passage?

Thanks!

hi,

As we know, there is a field named "model_input" in the test.json. It is very import for the generator model. But it is gone in the private test dataset. Can you give me some suggestions how to fill the ‘model_input’ field ? Thanks.

Hi @czyssrs, thanks for your excellent work

I only found these descriptions available in the README.file.

"pre_text": the texts before the table;
"post_text": the text after the table;
"table": the table;
"id": unique example id. composed by the original report name plus example index for this report. 

"qa": {
  "question": the question;
  "program": the reasoning program;
  "gold_inds": the gold supporting facts;
  "exe_ans": the gold execution result;
  "program_re": the reasoning program in nested format;
}

However, after walking through the actual train.json file, I found many other attributes like table_ori, table_retrieved, text_retrieved, etc. Just wondering if its possible that you could add a description for each attribute that in your dataset?

Also:

1. Isn't text_retrieved and table_retrieved with corresponding scores generated by a model?
2. what is model_input, tfidftopn under qa?
3. answer and exe_ans are different for Sample 1 (filename":"ADI/2009/page_49.pdf") in the train.json.
   • answer:380, exe_ans:3.8

Thanks for your help and support

def convert_single_mathqa_example(example, is_training, tokenizer, max_seq_length,
                                  max_program_length, op_list, op_list_size,
                                  const_list, const_list_size,
                                  cls_token, sep_token):
    """Converts a single MathQAExample into an InputFeature."""
    features = []
    question_tokens = example.question_tokens
    if len(question_tokens) > max_seq_length - 2:
        print("too long")
        question_tokens = question_tokens[:max_seq_length - 2]
    tokens = [cls_token] + question_tokens + [sep_token]         # 1. This line add [cls_token] at beginning.
    segment_ids = [0] * len(tokens)

    input_ids = tokenizer.convert_tokens_to_ids(tokens)

    input_mask = [1] * len(input_ids)
    for ind, offset in enumerate(example.number_indices):          # 2. Why don't number_indices offset by 1 ?
        if offset < len(input_mask):
            input_mask[offset] = 2
        else:
            if is_training == True:

                # invalid example, drop for training
                return features

            # assert is_training == False

Hello, Thanks for the great work! However, I am confused with the code. In the 1. comment, you add [cls_token] in front of the tokens, which means that the indices of tokens in the tokens will shift to the right by 1. In. 2. comment, you just use the example.number_indices to assign 2 to the indices of numbers, this is confusing, since input_mask is created from the tokens, which contains the [cls] at the beginning. For example: tokens: [[cls], a, b, 1, c, d], the example.number_indices will be [2] (because when you calculate the example.number_indices, there is no [cls] at the beginning, the "2" refers to the number "1"'s index ), the corresponding input_mask will be [1, 1, 1, 1, 1, 1]. When you try to assign the numbers' indices to 2 by the example.number_indices , the input_mask will be [1, 1, 0, 1, 1, 1], however, the 0'index 2 refers to the "b" in the tokens. Could you please explain this? Thanks!