====== Perplexity statistics ======
Mean PPL(Q)                   :   7.990955 ±   0.051527
Mean PPL(base)                :   7.534124 ±   0.048206
Cor(ln(PPL(Q)), ln(PPL(base))):  99.03%
Mean ln(PPL(Q)/PPL(base))     :   0.058868 ±   0.000897
Mean PPL(Q)/PPL(base)         :   1.060635 ±   0.000951
Mean PPL(Q)-PPL(base)         :   0.456831 ±   0.007702

====== KL divergence statistics ======
Mean    KLD:   0.051341 ±   0.000211
Maximum KLD:   6.532913
99.9%   KLD:   0.841224
99.0%   KLD:   0.299998
99.0%   KLD:   0.299998
Median  KLD:   0.038058
10.0%   KLD:   0.002190
 5.0%   KLD:   0.000690
 1.0%   KLD:   0.000095
Minimum KLD:  -0.000027

====== Token probability statistics ======
Mean    Δp: -1.276 ± 0.017 %
Maximum Δp: 97.255%
99.9%   Δp: 27.515%
99.0%   Δp: 15.140%
95.0%   Δp:  7.440%
90.0%   Δp:  4.054%
75.0%   Δp:  0.520%
Median  Δp: -0.129%
25.0%   Δp: -2.686%
10.0%   Δp: -8.215%
 5.0%   Δp: -12.358%
 1.0%   Δp: -22.757%
 0.1%   Δp: -44.994%
Minimum Δp: -85.147%
RMS Δp    :  6.512 ± 0.032 %
Same top p: 87.756 ± 0.086 %