====== Perplexity statistics ======
Mean PPL(Q)                   :  27.258871 ±0.268444
Mean PPL(base)                :  24.931431 ±0.241228
Cor(ln(PPL(Q)), ln(PPL(base))):  98.83%
Mean ln(PPL(Q)/PPL(base))     :   0.089250 ±0.001501
Mean PPL(Q)/PPL(base)         :   1.093354 ±0.001641
Mean PPL(Q)-PPL(base)         :   2.327441 ±0.047451

====== KL divergence statistics ======
Mean    KLD:   0.099355 ±0.000288
Maximum KLD:   3.087584
99.9%   KLD:   0.860643
99.0%   KLD:   0.485928
99.0%   KLD:   0.485928
Median  KLD:   0.067769
10.0%   KLD:   0.001206
 5.0%   KLD:   0.000193
 1.0%   KLD:   0.000008
Minimum KLD:  -0.000121

====== Token probability statistics ======
Mean    Δp: -1.067 ± 0.020 %
Maximum Δp: 80.056%
99.9%   Δp: 37.342%
99.0%   Δp: 21.343%
95.0%   Δp:  9.216%
90.0%   Δp:  4.374%
75.0%   Δp:  0.267%
Median  Δp: -0.010%
25.0%   Δp: -1.733%
10.0%   Δp: -8.613%
 5.0%   Δp: -14.630%
 1.0%   Δp: -27.490%
 0.1%   Δp: -44.569%
Minimum Δp: -75.190%
RMS Δp    :  7.613 ± 0.033 %
Same top p: 83.490 ± 0.096 %