The percentage of available discrete tokens in a codebook that are actually used during training or inference.