| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 16.574922 ± 0.145677 | |
| Mean PPL(base) : 13.052024 ± 0.108483 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 94.24% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.238948 ± 0.002939 | |
| Mean PPL(Q)/PPL(base) : 1.269912 ± 0.003733 | |
| Mean PPL(Q)-PPL(base) : 3.522897 ± 0.056595 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.367217 ± 0.001523 | |
| Maximum KLD: 17.225899 | |
| 99.9% KLD: 5.747202 | |
| 99.0% KLD: 2.751391 | |
| 99.0% KLD: 2.751391 | |
| Median KLD: 0.188572 | |
| 10.0% KLD: 0.004681 | |
| 5.0% KLD: 0.001022 | |
| 1.0% KLD: 0.000091 | |
| Minimum KLD: -0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.359 ± 0.041 % | |
| Maximum Δp: 98.460% | |
| 99.9% Δp: 72.658% | |
| 99.0% Δp: 39.222% | |
| 95.0% Δp: 18.089% | |
| 90.0% Δp: 9.430% | |
| 75.0% Δp: 0.936% | |
| Median Δp: -0.049% | |
| 25.0% Δp: -3.806% | |
| 10.0% Δp: -17.383% | |
| 5.0% Δp: -29.954% | |
| 1.0% Δp: -61.715% | |
| 0.1% Δp: -89.682% | |
| Minimum Δp: -99.948% | |
| RMS Δp : 15.613 ± 0.067 % | |
| Same top p: 74.188 ± 0.115 % | |