====== Perplexity statistics ======
Mean PPL(Q)                   :  10.063662 ±   0.073025
Mean PPL(base)                :  10.036835 ±   0.072696
Cor(ln(PPL(Q)), ln(PPL(base))):  99.97%
Mean ln(PPL(Q)/PPL(base))     :   0.002669 ±   0.000186
Mean PPL(Q)/PPL(base)         :   1.002673 ±   0.000186
Mean PPL(Q)-PPL(base)         :   0.026827 ±   0.001889

====== KL divergence statistics ======
Mean    KLD:   0.001360 ±   0.000005
Maximum KLD:   0.216463
99.9%   KLD:   0.021616
99.0%   KLD:   0.008655
99.0%   KLD:   0.008655
Median  KLD:   0.000893
10.0%   KLD:   0.000021
 5.0%   KLD:   0.000003
 1.0%   KLD:  -0.000002
Minimum KLD:  -0.000123

====== Token probability statistics ======
Mean    Δp: -0.011 ± 0.002 %
Maximum Δp: 26.248%
99.9%   Δp:  5.746%
99.0%   Δp:  2.880%
95.0%   Δp:  1.388%
90.0%   Δp:  0.805%
75.0%   Δp:  0.157%
Median  Δp: -0.000%
25.0%   Δp: -0.176%
10.0%   Δp: -0.830%
 5.0%   Δp: -1.418%
 1.0%   Δp: -2.952%
 0.1%   Δp: -6.178%
Minimum Δp: -16.853%
RMS Δp    :  0.953 ± 0.006 %
Same top p: 98.010 ± 0.036 %