====== Perplexity statistics ======
Mean PPL(Q)                   :   9.087639 ±   0.058224
Mean PPL(base)                :   7.534124 ±   0.048206
Cor(ln(PPL(Q)), ln(PPL(base))):  96.96%
Mean ln(PPL(Q)/PPL(base))     :   0.187473 ±   0.001579
Mean PPL(Q)/PPL(base)         :   1.206197 ±   0.001905
Mean PPL(Q)-PPL(base)         :   1.553515 ±   0.016466

====== KL divergence statistics ======
Mean    KLD:   0.173555 ±   0.000494
Maximum KLD:   5.439730
99.9%   KLD:   2.047228
99.0%   KLD:   0.821818
99.0%   KLD:   0.821818
Median  KLD:   0.143569
10.0%   KLD:   0.011327
 5.0%   KLD:   0.003726
 1.0%   KLD:   0.000528
Minimum KLD:   0.000000

====== Token probability statistics ======
Mean    Δp: -4.759 ± 0.031 %
Maximum Δp: 81.197%
99.9%   Δp: 38.726%
99.0%   Δp: 21.998%
95.0%   Δp:  9.357%
90.0%   Δp:  4.162%
75.0%   Δp:  0.081%
Median  Δp: -1.171%
25.0%   Δp: -8.442%
10.0%   Δp: -19.664%
 5.0%   Δp: -27.455%
 1.0%   Δp: -44.384%
 0.1%   Δp: -74.316%
Minimum Δp: -97.360%
RMS Δp    : 12.642 ± 0.047 %
Same top p: 77.787 ± 0.110 %