====== Perplexity statistics ======
Mean PPL(Q)                   :  10.774104 ±   0.078933
Mean PPL(base)                :  10.036835 ±   0.072696
Cor(ln(PPL(Q)), ln(PPL(base))):  98.73%
Mean ln(PPL(Q)/PPL(base))     :   0.070884 ±   0.001163
Mean PPL(Q)/PPL(base)         :   1.073456 ±   0.001249
Mean PPL(Q)-PPL(base)         :   0.737270 ±   0.013581

====== KL divergence statistics ======
Mean    KLD:   0.072166 ±   0.000301
Maximum KLD:   4.513918
99.9%   KLD:   1.255733
99.0%   KLD:   0.525244
99.0%   KLD:   0.525244
Median  KLD:   0.043659
10.0%   KLD:   0.001389
 5.0%   KLD:   0.000295
 1.0%   KLD:   0.000023
Minimum KLD:  -0.000033

====== Token probability statistics ======
Mean    Δp: -1.248 ± 0.018 %
Maximum Δp: 86.155%
99.9%   Δp: 34.118%
99.0%   Δp: 16.809%
95.0%   Δp:  7.276%
90.0%   Δp:  3.730%
75.0%   Δp:  0.431%
Median  Δp: -0.052%
25.0%   Δp: -2.217%
10.0%   Δp: -7.749%
 5.0%   Δp: -12.608%
 1.0%   Δp: -27.291%
 0.1%   Δp: -55.755%
Minimum Δp: -90.520%
RMS Δp    :  7.155 ± 0.041 %
Same top p: 86.887 ± 0.087 %