====== Perplexity statistics ======
Mean PPL(Q)                   :   8.320959 ±   0.058079
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  98.43%
Mean ln(PPL(Q)/PPL(base))     :   0.081564 ±   0.001230
Mean PPL(Q)/PPL(base)         :   1.084982 ±   0.001335
Mean PPL(Q)-PPL(base)         :   0.651747 ±   0.011213

====== KL divergence statistics ======
Mean    KLD:   0.083143 ±   0.000296
Maximum KLD:   6.062650
99.9%   KLD:   1.201464
99.0%   KLD:   0.480162
99.0%   KLD:   0.480162
Median  KLD:   0.056331
10.0%   KLD:   0.001167
 5.0%   KLD:   0.000247
 1.0%   KLD:   0.000025
Minimum KLD:  -0.000007

====== Token probability statistics ======
Mean    Δp: -1.276 ± 0.021 %
Maximum Δp: 67.711%
99.9%   Δp: 37.489%
99.0%   Δp: 21.789%
95.0%   Δp: 10.870%
90.0%   Δp:  5.938%
75.0%   Δp:  0.751%
Median  Δp: -0.048%
25.0%   Δp: -2.890%
10.0%   Δp: -10.428%
 5.0%   Δp: -16.263%
 1.0%   Δp: -29.041%
 0.1%   Δp: -51.259%
Minimum Δp: -95.041%
RMS Δp    :  8.464 ± 0.035 %
Same top p: 85.174 ± 0.091 %