====== Perplexity statistics ======
Mean PPL(Q)                   :  24.967196 ±   0.239198
Mean PPL(base)                :  22.656280 ±   0.216110
Cor(ln(PPL(Q)), ln(PPL(base))):  97.50%
Mean ln(PPL(Q)/PPL(base))     :   0.097126 ±   0.002137
Mean PPL(Q)/PPL(base)         :   1.101999 ±   0.002355
Mean PPL(Q)-PPL(base)         :   2.310915 ±   0.055814

====== KL divergence statistics ======
Mean    KLD:   0.194299 ±   0.000681
Maximum KLD:   7.302684
99.9%   KLD:   2.411403
99.0%   KLD:   1.223937
99.0%   KLD:   1.223937
Median  KLD:   0.115222
10.0%   KLD:   0.002397
 5.0%   KLD:   0.000468
 1.0%   KLD:   0.000029
Minimum KLD:  -0.000071

====== Token probability statistics ======
Mean    Δp: -1.814 ± 0.028 %
Maximum Δp: 87.005%
99.9%   Δp: 52.936%
99.0%   Δp: 28.197%
95.0%   Δp: 11.558%
90.0%   Δp:  5.344%
75.0%   Δp:  0.316%
Median  Δp: -0.027%
25.0%   Δp: -2.650%
10.0%   Δp: -12.214%
 5.0%   Δp: -20.885%
 1.0%   Δp: -41.395%
 0.1%   Δp: -70.258%
Minimum Δp: -96.511%
RMS Δp    : 10.877 ± 0.050 %
Same top p: 79.025 ± 0.105 %