Yes, diversity will be a problem, maybe I could look into the truffles for different tags. I have to see whether it even make sense to train different models on different tags (yet in this case the data might become quite sparse, especially for rare tags like pregnancy).