====== Perplexity statistics ======
Mean PPL(Q)                   :   7.895583 ±   0.054759
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  99.38%
Mean ln(PPL(Q)/PPL(base))     :   0.029090 ±   0.000771
Mean PPL(Q)/PPL(base)         :   1.029517 ±   0.000794
Mean PPL(Q)-PPL(base)         :   0.226371 ±   0.006353

====== KL divergence statistics ======
Mean    KLD:   0.031400 ±   0.000136
Maximum KLD:   6.726107
99.9%   KLD:   0.532302
99.0%   KLD:   0.204553
99.0%   KLD:   0.204553
Median  KLD:   0.020073
10.0%   KLD:   0.000422
 5.0%   KLD:   0.000093
 1.0%   KLD:   0.000009
Minimum KLD:  -0.000013

====== Token probability statistics ======
Mean    Δp: -0.223 ± 0.014 %
Maximum Δp: 66.844%
99.9%   Δp: 27.325%
99.0%   Δp: 15.333%
95.0%   Δp:  7.936%
90.0%   Δp:  4.574%
75.0%   Δp:  0.828%
Median  Δp: -0.003%
25.0%   Δp: -1.190%
10.0%   Δp: -5.309%
 5.0%   Δp: -8.723%
 1.0%   Δp: -16.788%
 0.1%   Δp: -33.037%
Minimum Δp: -87.879%
RMS Δp    :  5.278 ± 0.027 %
Same top p: 90.875 ± 0.074 %