====== Perplexity statistics ======
Mean PPL(Q)                   :   7.868503 ±   0.054510
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  99.45%
Mean ln(PPL(Q)/PPL(base))     :   0.025654 ±   0.000725
Mean PPL(Q)/PPL(base)         :   1.025986 ±   0.000744
Mean PPL(Q)-PPL(base)         :   0.199291 ±   0.005928

====== KL divergence statistics ======
Mean    KLD:   0.027967 ±   0.000110
Maximum KLD:   2.536634
99.9%   KLD:   0.471043
99.0%   KLD:   0.172088
99.0%   KLD:   0.172088
Median  KLD:   0.018325
10.0%   KLD:   0.000387
 5.0%   KLD:   0.000084
 1.0%   KLD:   0.000008
Minimum KLD:  -0.000068

====== Token probability statistics ======
Mean    Δp: -0.215 ± 0.013 %
Maximum Δp: 65.443%
99.9%   Δp: 25.656%
99.0%   Δp: 14.380%
95.0%   Δp:  7.485%
90.0%   Δp:  4.320%
75.0%   Δp:  0.792%
Median  Δp: -0.004%
25.0%   Δp: -1.139%
10.0%   Δp: -5.026%
 5.0%   Δp: -8.296%
 1.0%   Δp: -15.833%
 0.1%   Δp: -30.189%
Minimum Δp: -68.370%
RMS Δp    :  4.959 ± 0.024 %
Same top p: 91.376 ± 0.072 %