====== Perplexity statistics ======
Mean PPL(Q)                   :   8.536738 ±   0.059988
Mean PPL(base)                :   7.669212 ±   0.052592
Cor(ln(PPL(Q)), ln(PPL(base))):  98.03%
Mean ln(PPL(Q)/PPL(base))     :   0.107165 ±   0.001387
Mean PPL(Q)/PPL(base)         :   1.113118 ±   0.001544
Mean PPL(Q)-PPL(base)         :   0.867526 ±   0.013369

====== KL divergence statistics ======
Mean    KLD:   0.103999 ±   0.000378
Maximum KLD:   6.555358
99.9%   KLD:   1.574757
99.0%   KLD:   0.641801
99.0%   KLD:   0.641801
Median  KLD:   0.067611
10.0%   KLD:   0.001619
 5.0%   KLD:   0.000373
 1.0%   KLD:   0.000041
Minimum KLD:  -0.000000

====== Token probability statistics ======
Mean    Δp: -1.565 ± 0.023 %
Maximum Δp: 97.124%
99.9%   Δp: 40.899%
99.0%   Δp: 23.088%
95.0%   Δp: 11.311%
90.0%   Δp:  6.129%
75.0%   Δp:  0.713%
Median  Δp: -0.087%
25.0%   Δp: -3.327%
10.0%   Δp: -11.493%
 5.0%   Δp: -17.803%
 1.0%   Δp: -32.101%
 0.1%   Δp: -58.562%
Minimum Δp: -99.828%
RMS Δp    :  9.251 ± 0.039 %
Same top p: 84.006 ± 0.094 %