====== Perplexity statistics ======
Mean PPL(Q)                   :   6.494645 ±   0.044686
Mean PPL(base)                :   6.180577 ±   0.041038
Cor(ln(PPL(Q)), ln(PPL(base))):  99.22%
Mean ln(PPL(Q)/PPL(base))     :   0.049566 ±   0.000879
Mean PPL(Q)/PPL(base)         :   1.050815 ±   0.000924
Mean PPL(Q)-PPL(base)         :   0.314068 ±   0.006483

====== KL divergence statistics ======
Mean    KLD:   0.034702 ±   0.000219
Maximum KLD:   4.811197
99.9%   KLD:   1.115692
99.0%   KLD:   0.326864
99.0%   KLD:   0.326864
Median  KLD:   0.014286
10.0%   KLD:   0.000247
 5.0%   KLD:   0.000074
 1.0%   KLD:   0.000012
Minimum KLD:  -0.000329

====== Token probability statistics ======
Mean    Δp:  0.361 ± 0.015 %
Maximum Δp: 89.616%
99.9%   Δp: 34.967%
99.0%   Δp: 16.855%
95.0%   Δp:  8.285%
90.0%   Δp:  5.135%
75.0%   Δp:  1.359%
Median  Δp:  0.023%
25.0%   Δp: -0.451%
10.0%   Δp: -3.419%
 5.0%   Δp: -6.672%
 1.0%   Δp: -17.693%
 0.1%   Δp: -45.573%
Minimum Δp: -94.469%
RMS Δp    :  5.662 ± 0.041 %
Same top p: 92.468 ± 0.068 %