====== Perplexity statistics ======
Mean PPL(Q)                   :   7.815683 ±   0.054370
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  99.64%
Mean ln(PPL(Q)/PPL(base))     :   0.018918 ±   0.000594
Mean PPL(Q)/PPL(base)         :   1.019099 ±   0.000606
Mean PPL(Q)-PPL(base)         :   0.146471 ±   0.004873

====== KL divergence statistics ======
Mean    KLD:   0.015912 ±   0.000100
Maximum KLD:   7.292192
99.9%   KLD:   0.409283
99.0%   KLD:   0.134662
99.0%   KLD:   0.134662
Median  KLD:   0.007713
10.0%   KLD:   0.000195
 5.0%   KLD:   0.000042
 1.0%   KLD:   0.000003
Minimum KLD:  -0.000041

====== Token probability statistics ======
Mean    Δp:  0.068 ± 0.009 %
Maximum Δp: 66.453%
99.9%   Δp: 21.646%
99.0%   Δp: 10.534%
95.0%   Δp:  5.254%
90.0%   Δp:  3.187%
75.0%   Δp:  0.752%
Median  Δp:  0.001%
25.0%   Δp: -0.518%
10.0%   Δp: -2.750%
 5.0%   Δp: -4.853%
 1.0%   Δp: -11.600%
 0.1%   Δp: -27.299%
Minimum Δp: -69.966%
RMS Δp    :  3.642 ± 0.025 %
Same top p: 94.325 ± 0.059 %