Narrative QA Manual

[rsanjaykamath]

Submitting the manual version of Narrative QA script which requires a manual download from the original repository

[rsanjaykamath]

@lhoestq sorry I opened a new pull request because of some issues with the previous code base. This pull request is originally from #1364

[lhoestq]

Really cool thanks ! And good dataset card as well :)

I left a few comments

[rsanjaykamath]

Excellent comments. Thanks for those valuable suggestions. I changed everything as you have pointed out :)

[lhoestq]

Thanks for moving all the processing to _generate_examples !

I was wondering whether we should use the same features format as in the narrative_qa dataset script (same column names, same structure) for consistency. Let me know what you think

[rsanjaykamath]

I've copied the same template as NarrativeQA now. Please let me know if this is fine.

[lhoestq]

Awesome thank you !!
This looks all good :)

Just before we merge, I was wondering if you knew why the number of examples in the train set went from 1102 to 32747 in your last commit ? I can't see why the changes in the code would cause such a big difference

[rsanjaykamath]

Awesome thank you !!
This looks all good :)

Just before we merge, I was wondering if you knew why the number of examples in the train set went from 1102 to 32747 in your last commit ? I can't see why the changes in the code would cause such a big difference

Ok the change was the way I presented the data.
In my previous code, I presented a story with a list of questions-answers related to the story per sample. So the total 1102 was the number of stories (not questions) in the train set.

In the case of NarrativeQA, the code presented each sample data with one single question. So the story gets replicated as many times based on number of questions per story. I felt this was not really memory efficient so I had coded the way I did earlier.

But since this would be inconsistent as you pointed out, I modified my code to suit the NarrativeQA approach. Hope it's clear now :)

[lhoestq]

Ok I see ! that makes sense

[lhoestq]

Thanks for adding this one :)

[rsanjaykamath]

Thanks for your time and helping me with all this :) Really appreciate the hardwork you guys do.