cpral/nex-mix-17

-- A perplexity: 3.27196908 -- B perplexity: 3.29504072 -- A label in top-K: K = 1: 0.7130 K = 2: 0.8114 K = 3: 0.8531 K = 4: 0.8765 K = 5: 0.8925 -- B label in top-K: K = 1: 0.7113 K = 2: 0.8104 K = 3: 0.8525 K = 4: 0.8759 K = 5: 0.8923 -- Top-K agreement, A vs B: K = 1: 0.9524 K = 2: 0.8072 K = 3: 0.6241 K = 4: 0.4515 K = 5: 0.3110 -- KL divergence (A, B): 0.02734267 -- KL divergence (B, A): 0.02632093