====== Perplexity statistics ======
Mean PPL(Q)                   :   8.674113 ±   0.067361
Mean PPL(base)                :   8.445938 ±   0.065177
Cor(ln(PPL(Q)), ln(PPL(base))):  98.92%
Mean ln(PPL(Q)/PPL(base))     :   0.026657 ±   0.001137
Mean PPL(Q)/PPL(base)         :   1.027016 ±   0.001168
Mean PPL(Q)-PPL(base)         :   0.228175 ±   0.009969

====== KL divergence statistics ======
Mean    KLD:   0.043268 ±   0.000351
Maximum KLD:  18.135883
99.9%   KLD:   1.491586
99.0%   KLD:   0.417118
99.0%   KLD:   0.417118
Median  KLD:   0.016406
10.0%   KLD:   0.000034
 5.0%   KLD:   0.000004
 1.0%   KLD:  -0.000000
Minimum KLD:  -0.000026

====== Token probability statistics ======
Mean    Δp: -0.672 ± 0.017 %
Maximum Δp: 98.835%
99.9%   Δp: 40.761%
99.0%   Δp: 17.002%
95.0%   Δp:  6.983%
90.0%   Δp:  3.441%
75.0%   Δp:  0.273%
Median  Δp: -0.003%
25.0%   Δp: -1.108%
10.0%   Δp: -5.679%
 5.0%   Δp: -10.036%
 1.0%   Δp: -23.278%
 0.1%   Δp: -54.994%
Minimum Δp: -99.427%
RMS Δp    :  6.596 ± 0.049 %
Same top p: 91.212 ± 0.073 %