Cool dataset!
Congrats, this looks cool! Large-scale fiction datasets are always super neat, so props on taking the time to upload this.
Saw your reddit thread, but it's filled with people who are being negative for no real reason and I thought it might be better to post here. (I don't know what it is about reddit that turns people into jerks and naysayers...)
These are just some comments that occured to me as somebody who's worked on similar datasets.
1.) Just glancing at it there appears to be some oddball stuff that might be a scraping issue. For instance, your 9th row or so starts with "Ch Child Compensation" in the English even though that doesn't appear in the Japanese text. I haven't gone through it thoroughly but that jumped out at me.
2.) It feels like the dataset could benefit from more metadata than you've currently got. For instance, Japanese story title, genre, translator name (If available? I don't read scraper sites) so I don't know what's there), maybe if the original source was narou/kakuyomu/etc. since those tend to have varied levels of quality... If there's any kind of quality ratings available, etc.
The kind of metadata you want is obviously going to depend on intended use case but it feels like there could be more there.
All in all very cool, though, and look forward to seeing what you do with it.
Those are a lot of good points,
For #1, that's hard to fix, as the sentence aligner splits the documents at line breaks and then matches them, so if there's an addition to a line but most of it matches, it'll get passed. The only way to fix that is to change the alignment threshold, but that would decrease the number of less direct translations that get passed significantly, another downside. I looked through the dataset, and examples like that are thankfully rare enough to hopefully not pose issues.
For #2, that's something I hadn't thought of, and should be relatively easy to fix. I'll look into it and see what I can add.