That's funny, I did the same analysis with sci-hub. Back when there was an organized drive to back it up.

I downloaded parts of it and wanted to figure out why it was so heavy, seeing as you'd expect articles to be mostly text and very light.

There was a similar distribution of file sizes. My immediate instinct was also to cut off the tail-end, but looking at the larger files I realized it was a whole range of good articles that included high quality graphics that were crucial to the research being presented, not poor compression or useless bloat.