Could split the stories into buckets and then randomly sample from each bucket. Most stories are small, so they’re currently overrepresented in the sampling.