====== Perplexity statistics ======
Mean PPL(Q)                   :   7.347298 ±   0.046298
Mean PPL(base)                :   7.237090 ±   0.045539
Cor(ln(PPL(Q)), ln(PPL(base))):  99.80%
Mean ln(PPL(Q)/PPL(base))     :   0.015114 ±   0.000397
Mean PPL(Q)/PPL(base)         :   1.015228 ±   0.000403
Mean PPL(Q)-PPL(base)         :   0.110209 ±   0.002994

====== KL divergence statistics ======
Mean    KLD:   0.009763 ±   0.000056
Maximum KLD:   1.979661
99.9%   KLD:   0.248941
99.0%   KLD:   0.074428
99.0%   KLD:   0.074428
Median  KLD:   0.005853
10.0%   KLD:   0.000390
 5.0%   KLD:   0.000129
 1.0%   KLD:   0.000017
Minimum KLD:  -0.000066

====== Token probability statistics ======
Mean    Δp: -0.342 ± 0.007 %
Maximum Δp: 55.623%
99.9%   Δp: 13.960%
99.0%   Δp:  6.651%
95.0%   Δp:  3.167%
90.0%   Δp:  1.790%
75.0%   Δp:  0.310%
Median  Δp: -0.024%
25.0%   Δp: -0.832%
10.0%   Δp: -2.797%
 5.0%   Δp: -4.458%
 1.0%   Δp: -9.272%
 0.1%   Δp: -21.545%
Minimum Δp: -83.712%
RMS Δp    :  2.763 ± 0.024 %
Same top p: 95.200 ± 0.056 %