====== Perplexity statistics ====== | |
Mean PPL(Q) : 10.469704 ± 0.076871 | |
Mean PPL(base) : 10.036835 ± 0.072696 | |
Cor(ln(PPL(Q)), ln(PPL(base))): 99.20% | |
Mean ln(PPL(Q)/PPL(base)) : 0.042224 ± 0.000927 | |
Mean PPL(Q)/PPL(base) : 1.043128 ± 0.000967 | |
Mean PPL(Q)-PPL(base) : 0.432869 ± 0.010328 | |
====== KL divergence statistics ====== | |
Mean KLD: 0.044002 ± 0.000201 | |
Maximum KLD: 5.247561 | |
99.9% KLD: 0.866438 | |
99.0% KLD: 0.326842 | |
99.0% KLD: 0.326842 | |
Median KLD: 0.025367 | |
10.0% KLD: 0.000731 | |
5.0% KLD: 0.000142 | |
1.0% KLD: 0.000009 | |
Minimum KLD: -0.000234 | |
====== Token probability statistics ====== | |
Mean Δp: -0.465 ± 0.014 % | |
Maximum Δp: 82.693% | |
99.9% Δp: 28.711% | |
99.0% Δp: 14.576% | |
95.0% Δp: 6.726% | |
90.0% Δp: 3.801% | |
75.0% Δp: 0.673% | |
Median Δp: -0.005% | |
25.0% Δp: -1.170% | |
10.0% Δp: -5.104% | |
5.0% Δp: -8.747% | |
1.0% Δp: -19.949% | |
0.1% Δp: -42.515% | |
Minimum Δp: -96.762% | |
RMS Δp : 5.550 ± 0.036 % | |
Same top p: 89.641 ± 0.079 % | |