This is fantastic, @mulhod! Thanks for taking this on – this is a really nice improvement of the codebase.

I made a couple of minor suggestions. I also have a couple of broader suggestions:

1. Can you run the tests locally to make sure that all the created files are actually being cleaned up properly? As long as we are working on this, it would be good to make sure of those.
2. I am a little concerned about changing the from tests.utils import X statements to from .utils import X in terms of the impact this will have on the testing of the PyPi and conda packages. In those scenarios, we install the conda package, do a sparse checkout of the skll tests and then run the tests. I think you should be able to test this by making a local conda package and simulating this scenario.

Thanks again!

Thanks, @desilinguist! I decided to revert the parts that used relative imports for now. It's a minor change, anyway. I have run all the tests and fixed a couple cases (one that you pointed out!). I think most of the cases were actually issues with the tests that didn't have anything to do with my changes. But, this gave me the chance to really see what files are left behind, so I just fixed any I saw.