====== Perplexity statistics ======
Mean PPL(Q)                   :   8.950671 ±   0.069566
Mean PPL(base)                :   8.445938 ±   0.065177
Cor(ln(PPL(Q)), ln(PPL(base))):  97.80%
Mean ln(PPL(Q)/PPL(base))     :   0.058043 ±   0.001625
Mean PPL(Q)/PPL(base)         :   1.059760 ±   0.001722
Mean PPL(Q)-PPL(base)         :   0.504733 ±   0.014789

====== KL divergence statistics ======
Mean    KLD:   0.095434 ±   0.000689
Maximum KLD:  32.811981
99.9%   KLD:   2.871047
99.0%   KLD:   0.908818
99.0%   KLD:   0.908818
Median  KLD:   0.038622
10.0%   KLD:   0.000091
 5.0%   KLD:   0.000010
 1.0%   KLD:   0.000000
Minimum KLD:  -0.000006

====== Token probability statistics ======
Mean    Δp: -1.299 ± 0.025 %
Maximum Δp: 99.105%
99.9%   Δp: 51.217%
99.0%   Δp: 24.267%
95.0%   Δp: 10.059%
90.0%   Δp:  4.891%
75.0%   Δp:  0.312%
Median  Δp: -0.013%
25.0%   Δp: -1.967%
10.0%   Δp: -9.334%
 5.0%   Δp: -16.184%
 1.0%   Δp: -35.890%
 0.1%   Δp: -76.357%
Minimum Δp: -99.865%
RMS Δp    :  9.560 ± 0.056 %
Same top p: 87.237 ± 0.086 %