How do you evaluate the performance of a model when ground-truth labels are delayed or sparse on Twitch?
