Each dataset have its own license, see the readme/license for details.

The datasets can be inspected here: ARC-Interactive

Ideally there should be no suffixes. The _v2 and _v3 suffixes are my kludgy workaround. However when recording interaction histories, then one interaction history file references a specific version of a task. When a task gets fixed upstream, then the old solution may no longer be the correct solution for the new task. And deleting old interaction history files that would be sad. So _v it was.

If there is a _v suffix then the newest is the highest version.

I also considered going with commit ids, but that would have made it even more tricky to keep track of.