====== Perplexity statistics ======
Mean PPL(Q)                   :  10.071679 ±   0.073178
Mean PPL(base)                :  10.036835 ±   0.072696
Cor(ln(PPL(Q)), ln(PPL(base))):  99.93%
Mean ln(PPL(Q)/PPL(base))     :   0.003466 ±   0.000272
Mean PPL(Q)/PPL(base)         :   1.003472 ±   0.000273
Mean PPL(Q)-PPL(base)         :   0.034844 ±   0.002769

====== KL divergence statistics ======
Mean    KLD:   0.003501 ±   0.000018
Maximum KLD:   1.326304
99.9%   KLD:   0.063485
99.0%   KLD:   0.024422
99.0%   KLD:   0.024422
Median  KLD:   0.002144
10.0%   KLD:   0.000053
 5.0%   KLD:   0.000008
 1.0%   KLD:  -0.000001
Minimum KLD:  -0.000235

====== Token probability statistics ======
Mean    Δp:  0.013 ± 0.004 %
Maximum Δp: 34.867%
99.9%   Δp:  9.405%
99.0%   Δp:  4.684%
95.0%   Δp:  2.228%
90.0%   Δp:  1.306%
75.0%   Δp:  0.272%
Median  Δp:  0.000%
25.0%   Δp: -0.249%
10.0%   Δp: -1.227%
 5.0%   Δp: -2.154%
 1.0%   Δp: -4.685%
 0.1%   Δp: -10.648%
Minimum Δp: -30.875%
RMS Δp    :  1.532 ± 0.011 %
Same top p: 96.908 ± 0.045 %