It's a shame that Github's large file storage (LFS) mechanism isn't publicly released yet, because I am sure that will appease your problems. (This is assuming your datasets don't fit into a csv (note that the current hard limit for github files is 100MB, and a 100MB CSV file is a lot of data)).

I was thinking more along the lines of data that you'd want to graph or preset for analysis, as in a notebook or a paper. I would (hopefully) assume that your simulation data would be generated and used in such a way that it wouldn't need to be stored permanently. At least, when I was doing simulation based analysis, I wasn't necessarily concerned about any individual run, but rather a combination of a bunch of simulations (all of which were ephemeral).

Nor, to be blunt, should they.

My particular simulation work is rather interested both in individual runs (and indeed, individuals within those runs) as well as summarization.

Beyond that, what use is there to putting just the "summary" data online, when the underlying data made that still exists as "you're just going to have to trust me". Being able to replicate my figure code doesn't get people very far.

There's also a weird data rabbit hole. Is the code that generates the simulation data enough, without the underlying data that code uses? Some of that is either protected or proprietary, so even with the code, it's utterly useless.

Surely, it does not help with private or mutable data, but how would magnet URIs and bittorrent help? If it is private, it cannot be shared with bittorrent either and sharing mutable data via bittorrent is not its primary use case.

Bittorrent might help with big data, where big means 100MB+ in the case of Github. There are other approaches for big files in git (git annex, Github's LFS, git-bigfiles, etc).