Run no-hot-spots-in-training

Test hypothesis that training ML models without hot spots will allow ML model to make better predictions on cold spots (i.e. ML models like one mean value) while also better predict which sites are dramatically different (i.e. the hot spots) thus helping identify them. Otherwise, ML models tend to gravitate to the mean, which is right in between the cold and hot spots in the "valley of IDK" of the bimodal distribution.