Maybe we have talked about this before but I forgot. Should we use human label or machine label, or both of them for training? Human labels have more unique labels than machine label (18155 vs. 7509). But machine labels have less confidence (say, confidence might be 0.7 instead of 1 in human label.) Moreover, more labels in human label file are not trainable. (there are only 7178 trainable label). But, in total there are fewer human labels. (8036466 vs 15259186), so from the perspective of efficiency and convenience, we should probably start with human label and get its subset of "trainable" ones?