====== Perplexity statistics ======
Mean PPL(Q)                   :  10.092163 ±   0.073418
Mean PPL(base)                :  10.036835 ±   0.072696
Cor(ln(PPL(Q)), ln(PPL(base))):  99.92%
Mean ln(PPL(Q)/PPL(base))     :   0.005497 ±   0.000290
Mean PPL(Q)/PPL(base)         :   1.005513 ±   0.000291
Mean PPL(Q)-PPL(base)         :   0.055329 ±   0.002985

====== KL divergence statistics ======
Mean    KLD:   0.004008 ±   0.000021
Maximum KLD:   1.392368
99.9%   KLD:   0.076297
99.0%   KLD:   0.029280
99.0%   KLD:   0.029280
Median  KLD:   0.002400
10.0%   KLD:   0.000055
 5.0%   KLD:   0.000008
 1.0%   KLD:  -0.000002
Minimum KLD:  -0.000235

====== Token probability statistics ======
Mean    Δp:  0.003 ± 0.004 %
Maximum Δp: 44.164%
99.9%   Δp: 10.258%
99.0%   Δp:  4.882%
95.0%   Δp:  2.317%
90.0%   Δp:  1.357%
75.0%   Δp:  0.286%
Median  Δp: -0.000%
25.0%   Δp: -0.264%
10.0%   Δp: -1.302%
 5.0%   Δp: -2.305%
 1.0%   Δp: -5.091%
 0.1%   Δp: -11.711%
Minimum Δp: -28.870%
RMS Δp    :  1.636 ± 0.011 %
Same top p: 96.720 ± 0.046 %