| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 33.128433 ±0.335060 | |
| Mean PPL(base) : 24.931431 ±0.241228 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 96.34% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.284263 ±0.002714 | |
| Mean PPL(Q)/PPL(base) : 1.328782 ±0.003606 | |
| Mean PPL(Q)-PPL(base) : 8.197002 ±0.121360 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.345716 ±0.000950 | |
| Maximum KLD: 6.695283 | |
| 99.9% KLD: 2.655785 | |
| 99.0% KLD: 1.596549 | |
| 99.0% KLD: 1.596549 | |
| Median KLD: 0.245697 | |
| 10.0% KLD: 0.005232 | |
| 5.0% KLD: 0.000912 | |
| 1.0% KLD: 0.000055 | |
| Minimum KLD: -0.000028 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.627 ± 0.034 % | |
| Maximum Δp: 96.502% | |
| 99.9% Δp: 56.095% | |
| 99.0% Δp: 33.635% | |
| 95.0% Δp: 14.828% | |
| 90.0% Δp: 6.605% | |
| 75.0% Δp: 0.200% | |
| Median Δp: -0.076% | |
| 25.0% Δp: -3.969% | |
| 10.0% Δp: -16.576% | |
| 5.0% Δp: -27.387% | |
| 1.0% Δp: -50.164% | |
| 0.1% Δp: -74.261% | |
| Minimum Δp: -95.075% | |
| RMS Δp : 13.304 ± 0.052 % | |
| Same top p: 72.106 ± 0.116 % | |