| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.955300 ± 0.051312 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.11% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.054396 ± 0.000860 | |
| Mean PPL(Q)/PPL(base) : 1.055902 ± 0.000908 | |
| Mean PPL(Q)-PPL(base) : 0.421176 ± 0.007333 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.043200 ± 0.000262 | |
| Maximum KLD: 8.829779 | |
| 99.9% KLD: 1.281964 | |
| 99.0% KLD: 0.376543 | |
| 99.0% KLD: 0.376543 | |
| Median KLD: 0.023787 | |
| 10.0% KLD: 0.001665 | |
| 5.0% KLD: 0.000574 | |
| 1.0% KLD: 0.000087 | |
| Minimum KLD: -0.000360 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -1.119 ± 0.016 % | |
| Maximum Δp: 59.041% | |
| 99.9% Δp: 24.874% | |
| 99.0% Δp: 12.145% | |
| 95.0% Δp: 5.639% | |
| 90.0% Δp: 3.126% | |
| 75.0% Δp: 0.437% | |
| Median Δp: -0.107% | |
| 25.0% Δp: -1.966% | |
| 10.0% Δp: -6.121% | |
| 5.0% Δp: -9.877% | |
| 1.0% Δp: -23.039% | |
| 0.1% Δp: -56.479% | |
| Minimum Δp: -96.588% | |
| RMS Δp : 6.074 ± 0.047 % | |
| Same top p: 90.347 ± 0.078 % | |