====== Perplexity statistics ======
Mean PPL(Q)                   :   9.058327 ±   0.069783
Mean PPL(base)                :   8.445938 ±   0.065177
Cor(ln(PPL(Q)), ln(PPL(base))):  97.14%
Mean ln(PPL(Q)/PPL(base))     :   0.069999 ±   0.001845
Mean PPL(Q)/PPL(base)         :   1.072507 ±   0.001979
Mean PPL(Q)-PPL(base)         :   0.612389 ±   0.016784

====== KL divergence statistics ======
Mean    KLD:   0.127036 ±   0.000807
Maximum KLD:  23.618629
99.9%   KLD:   3.421933
99.0%   KLD:   1.181114
99.0%   KLD:   1.181114
Median  KLD:   0.052644
10.0%   KLD:   0.000185
 5.0%   KLD:   0.000024
 1.0%   KLD:   0.000000
Minimum KLD:  -0.000005

====== Token probability statistics ======
Mean    Δp: -2.079 ± 0.028 %
Maximum Δp: 99.099%
99.9%   Δp: 55.973%
99.0%   Δp: 25.713%
95.0%   Δp: 10.217%
90.0%   Δp:  4.657%
75.0%   Δp:  0.194%
Median  Δp: -0.050%
25.0%   Δp: -2.870%
10.0%   Δp: -11.859%
 5.0%   Δp: -20.061%
 1.0%   Δp: -44.003%
 0.1%   Δp: -82.841%
Minimum Δp: -99.876%
RMS Δp    : 11.057 ± 0.060 %
Same top p: 85.550 ± 0.091 %