i would love to know more about this. some of my ideas are super simple dataset that might very well overfit, but show that the code works. other than that, i would love to hear more about simple tests which do not need to run the full pipeline on the full dataset (which does not tell you that much ultimately)