====== Perplexity statistics ======
Mean PPL(Q)                   :   6.557455 ±   0.040211
Mean PPL(base)                :   6.554978 ±   0.040159
Cor(ln(PPL(Q)), ln(PPL(base))):  99.96%
Mean ln(PPL(Q)/PPL(base))     :   0.000378 ±   0.000173
Mean PPL(Q)/PPL(base)         :   1.000378 ±   0.000173
Mean PPL(Q)-PPL(base)         :   0.002477 ±   0.001132

====== KL divergence statistics ======
Mean    KLD:   0.001610 ±   0.000012
Maximum KLD:   1.216474
99.9%   KLD:   0.035302
99.0%   KLD:   0.010267
99.0%   KLD:   0.010267
Median  KLD:   0.001088
10.0%   KLD:   0.000043
 5.0%   KLD:   0.000011
 1.0%   KLD:   0.000001
Minimum KLD:  -0.000107

====== Token probability statistics ======
Mean    Δp: -0.013 ± 0.003 %
Maximum Δp: 55.677%
99.9%   Δp:  7.060%
99.0%   Δp:  3.700%
95.0%   Δp:  1.915%
90.0%   Δp:  1.170%
75.0%   Δp:  0.266%
Median  Δp: -0.000%
25.0%   Δp: -0.285%
10.0%   Δp: -1.201%
 5.0%   Δp: -1.965%
 1.0%   Δp: -3.786%
 0.1%   Δp: -7.655%
Minimum Δp: -31.459%
RMS Δp    :  1.279 ± 0.012 %
Same top p: 97.555 ± 0.041 %