====== Perplexity statistics ======
Mean PPL(Q)                   :   7.570386 ±   0.047455
Mean PPL(base)                :   7.237090 ±   0.045539
Cor(ln(PPL(Q)), ln(PPL(base))):  99.27%
Mean ln(PPL(Q)/PPL(base))     :   0.045025 ±   0.000759
Mean PPL(Q)/PPL(base)         :   1.046054 ±   0.000794
Mean PPL(Q)-PPL(base)         :   0.333296 ±   0.005934

====== KL divergence statistics ======
Mean    KLD:   0.036155 ±   0.000172
Maximum KLD:   3.950742
99.9%   KLD:   0.799789
99.0%   KLD:   0.253352
99.0%   KLD:   0.253352
Median  KLD:   0.023291
10.0%   KLD:   0.001614
 5.0%   KLD:   0.000535
 1.0%   KLD:   0.000079
Minimum KLD:  -0.000198

====== Token probability statistics ======
Mean    Δp: -1.387 ± 0.014 %
Maximum Δp: 59.331%
99.9%   Δp: 22.210%
99.0%   Δp: 10.808%
95.0%   Δp:  4.768%
90.0%   Δp:  2.461%
75.0%   Δp:  0.216%
Median  Δp: -0.219%
25.0%   Δp: -2.513%
10.0%   Δp: -6.790%
 5.0%   Δp: -10.102%
 1.0%   Δp: -19.543%
 0.1%   Δp: -42.826%
Minimum Δp: -87.885%
RMS Δp    :  5.421 ± 0.034 %
Same top p: 90.686 ± 0.077 %