====== Perplexity statistics ======
Mean PPL(Q)                   :   8.286801 ±   0.058783
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  98.72%
Mean ln(PPL(Q)/PPL(base))     :   0.077450 ±   0.001141
Mean PPL(Q)/PPL(base)         :   1.080528 ±   0.001233
Mean PPL(Q)-PPL(base)         :   0.617589 ±   0.010843

====== KL divergence statistics ======
Mean    KLD:   0.061737 ±   0.000316
Maximum KLD:   6.420732
99.9%   KLD:   1.478346
99.0%   KLD:   0.532294
99.0%   KLD:   0.532294
Median  KLD:   0.029140
10.0%   KLD:   0.000817
 5.0%   KLD:   0.000195
 1.0%   KLD:   0.000021
Minimum KLD:  -0.000001

====== Token probability statistics ======
Mean    Δp: -0.421 ± 0.018 %
Maximum Δp: 96.151%
99.9%   Δp: 38.523%
99.0%   Δp: 18.770%
95.0%   Δp:  8.780%
90.0%   Δp:  5.065%
75.0%   Δp:  0.992%
Median  Δp: -0.007%
25.0%   Δp: -1.428%
10.0%   Δp: -6.125%
 5.0%   Δp: -10.610%
 1.0%   Δp: -24.391%
 0.1%   Δp: -54.028%
Minimum Δp: -92.165%
RMS Δp    :  6.903 ± 0.041 %
Same top p: 89.261 ± 0.080 %