As the title says, I'm working on building a logistic regression model. I've set up a function to flag outlier observations based on Tukey's fences, as a couple of my predictor variables are heavily skewed. I'm doing a 70/30 train/test split. Of 299 observations, 107 lie outside Tukey's fences.

What I'm wondering is this: Do I violate any conventions by also splitting the outliers 70/30? So my training set would consist of 70% of the "typical" observations and 70% of the outliers, and my testing set would be the other 30% of each. The thought process here is to build the model with adequate exposure to extreme cases, and also be able to test on those extreme cases. Thanks in advance!