Views
No views yet
eval/model_diff.py -- original perplexity: 3.18951417
-- original label in top-K:
K = 1: 0.7194
K = 2: 0.8281
K = 3: 0.8719
K = 4: 0.8930
K = 5: 0.9078
-- 5.0bpw-h6 perplexity: 3.21707290
-- 5.0bpw-h6 label in top-K:
K = 1: 0.7197
K = 2: 0.8254
K = 3: 0.8702
K = 4: 0.8921
K = 5: 0.9074
-- Top-K agreement, 5.0bpw-h6 vs original:
K = 1: 0.9487
K = 2: 0.8222
K = 3: 0.6600
K = 4: 0.4920
K = 5: 0.3487
-- KL divergence (A, original): 0.03612041