====== Perplexity statistics ======
Mean PPL(Q)                   :   8.459379 ±   0.053550
Mean PPL(base)                :   7.237090 ±   0.045539
Cor(ln(PPL(Q)), ln(PPL(base))):  97.26%
Mean ln(PPL(Q)/PPL(base))     :   0.156057 ±   0.001477
Mean PPL(Q)/PPL(base)         :   1.168892 ±   0.001727
Mean PPL(Q)-PPL(base)         :   1.222289 ±   0.014061

====== KL divergence statistics ======
Mean    KLD:   0.131196 ±   0.000539
Maximum KLD:   7.898368
99.9%   KLD:   2.475934
99.0%   KLD:   0.894390
99.0%   KLD:   0.894390
Median  KLD:   0.089346
10.0%   KLD:   0.006670
 5.0%   KLD:   0.002250
 1.0%   KLD:   0.000392
Minimum KLD:   0.000001

====== Token probability statistics ======
Mean    Δp: -3.913 ± 0.027 %
Maximum Δp: 64.023%
99.9%   Δp: 33.075%
99.0%   Δp: 17.214%
95.0%   Δp:  7.245%
90.0%   Δp:  3.301%
75.0%   Δp:  0.096%
Median  Δp: -0.875%
25.0%   Δp: -6.342%
10.0%   Δp: -15.578%
 5.0%   Δp: -22.649%
 1.0%   Δp: -41.943%
 0.1%   Δp: -75.852%
Minimum Δp: -98.926%
RMS Δp    : 10.892 ± 0.050 %
Same top p: 82.438 ± 0.100 %