Image files missing in mPLUG/MP-DocStruct1M

[ymzhang0303]

Thanks for your work! I found that there are many image files missing, for example, "pdfa-eng-wds/pdfa-eng-train-0034-pdfplumber/pages/6551295_page3.png" is used in multi_page_parsing.jsonl, but was not found in imgs.zip.

[HAWLYQ]

Hi, @ymzhang0303 Thanks for your reminder, we found just the first image of each sample is uploaded. We will update the images soon~

[ymzhang0303]

Thanks! Looking forward to it!

[ymzhang0303]

Also I have a question for processing PDFA dataset, when there are two columns in one page, looks like you are just connect the text in the same line from left to right, which is not a human-reading style. May I know the reason you do that? Thanks!

[HAWLYQ]

Hi, @ymzhang0303, images of MP-DocStruct1M have been updated on both HuggingFace and ModelScope~

[HAWLYQ]

Also I have a question for processing PDFA dataset, when there are two columns in one page, looks like you are just connect the text in the same line from left to right, which is not a human-reading style. May I know the reason you do that? Thanks!

Organizing texts in the reading order is indeed better, however, we didn't find a satisfying tool or human annotations to construct such training samples~