We set up dataset splits and the usual best practices. Of course, if you overdo things, you can still hack benchmarks; our goal isn't to publish SOTA numbers bu

We set up dataset splits and the usual best practices. Of course, if you overdo things, you can still hack benchmarks; our goal isn't to publish SOTA numbers but rather to illustrate results from our methodology. We didn't even tune hyperparameters, we just used the default choices. Definitely a valid concern for teams chasing SOTA though. Thanks!