====== Perplexity statistics ======
Mean PPL(Q)                   :   6.754714 ±   0.041611
Mean PPL(base)                :   6.554978 ±   0.040159
Cor(ln(PPL(Q)), ln(PPL(base))):  99.29%
Mean ln(PPL(Q)/PPL(base))     :   0.030016 ±   0.000731
Mean PPL(Q)/PPL(base)         :   1.030471 ±   0.000753
Mean PPL(Q)-PPL(base)         :   0.199735 ±   0.005068

====== KL divergence statistics ======
Mean    KLD:   0.033124 ±   0.000235
Maximum KLD:   6.718605
99.9%   KLD:   1.106657
99.0%   KLD:   0.315475
99.0%   KLD:   0.315475
Median  KLD:   0.016259
10.0%   KLD:   0.000792
 5.0%   KLD:   0.000230
 1.0%   KLD:   0.000033
Minimum KLD:  -0.000025

====== Token probability statistics ======
Mean    Δp: -0.664 ± 0.014 %
Maximum Δp: 95.789%
99.9%   Δp: 26.631%
99.0%   Δp: 11.752%
95.0%   Δp:  5.423%
90.0%   Δp:  3.151%
75.0%   Δp:  0.595%
Median  Δp: -0.035%
25.0%   Δp: -1.345%
10.0%   Δp: -4.724%
 5.0%   Δp: -7.884%
 1.0%   Δp: -20.310%
 0.1%   Δp: -52.412%
Minimum Δp: -97.482%
RMS Δp    :  5.475 ± 0.047 %
Same top p: 91.545 ± 0.073 %