====== Perplexity statistics ======
Mean PPL(Q)                   :   7.337057 ±   0.046220
Mean PPL(base)                :   7.237090 ±   0.045539
Cor(ln(PPL(Q)), ln(PPL(base))):  99.81%
Mean ln(PPL(Q)/PPL(base))     :   0.013719 ±   0.000386
Mean PPL(Q)/PPL(base)         :   1.013813 ±   0.000391
Mean PPL(Q)-PPL(base)         :   0.099967 ±   0.002893

====== KL divergence statistics ======
Mean    KLD:   0.009155 ±   0.000052
Maximum KLD:   2.029761
99.9%   KLD:   0.235197
99.0%   KLD:   0.070048
99.0%   KLD:   0.070048
Median  KLD:   0.005494
10.0%   KLD:   0.000351
 5.0%   KLD:   0.000113
 1.0%   KLD:   0.000014
Minimum KLD:  -0.000066

====== Token probability statistics ======
Mean    Δp: -0.318 ± 0.007 %
Maximum Δp: 56.530%
99.9%   Δp: 14.003%
99.0%   Δp:  6.553%
95.0%   Δp:  3.083%
90.0%   Δp:  1.762%
75.0%   Δp:  0.311%
Median  Δp: -0.019%
25.0%   Δp: -0.801%
10.0%   Δp: -2.703%
 5.0%   Δp: -4.299%
 1.0%   Δp: -8.935%
 0.1%   Δp: -20.874%
Minimum Δp: -84.208%
RMS Δp    :  2.680 ± 0.024 %
Same top p: 95.302 ± 0.056 %