====== Perplexity statistics ======
Mean PPL(Q)                   :   8.389337 ±   0.058940
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  98.30%
Mean ln(PPL(Q)/PPL(base))     :   0.089748 ±   0.001293
Mean PPL(Q)/PPL(base)         :   1.093898 ±   0.001414
Mean PPL(Q)-PPL(base)         :   0.720125 ±   0.012082

====== KL divergence statistics ======
Mean    KLD:   0.089559 ±   0.000330
Maximum KLD:   5.434262
99.9%   KLD:   1.383622
99.0%   KLD:   0.545214
99.0%   KLD:   0.545214
Median  KLD:   0.059201
10.0%   KLD:   0.001248
 5.0%   KLD:   0.000276
 1.0%   KLD:   0.000028
Minimum KLD:  -0.000002

====== Token probability statistics ======
Mean    Δp: -1.135 ± 0.022 %
Maximum Δp: 74.896%
99.9%   Δp: 38.879%
99.0%   Δp: 23.059%
95.0%   Δp: 11.570%
90.0%   Δp:  6.403%
75.0%   Δp:  0.894%
Median  Δp: -0.038%
25.0%   Δp: -2.772%
10.0%   Δp: -10.283%
 5.0%   Δp: -16.115%
 1.0%   Δp: -29.853%
 0.1%   Δp: -55.436%
Minimum Δp: -97.449%
RMS Δp    :  8.734 ± 0.037 %
Same top p: 84.881 ± 0.092 %