| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 9.016990 ± 0.059441 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 97.13% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.179668 ± 0.001569 | |
| Mean PPL(Q)/PPL(base) : 1.196820 ± 0.001878 | |
| Mean PPL(Q)-PPL(base) : 1.482867 ± 0.017057 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.157666 ± 0.000482 | |
| Maximum KLD: 9.472803 | |
| 99.9% KLD: 2.049311 | |
| 99.0% KLD: 0.765471 | |
| 99.0% KLD: 0.765471 | |
| Median KLD: 0.127495 | |
| 10.0% KLD: 0.007270 | |
| 5.0% KLD: 0.002221 | |
| 1.0% KLD: 0.000296 | |
| Minimum KLD: -0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -3.373 ± 0.030 % | |
| Maximum Δp: 68.206% | |
| 99.9% Δp: 40.367% | |
| 99.0% Δp: 24.601% | |
| 95.0% Δp: 11.563% | |
| 90.0% Δp: 5.922% | |
| 75.0% Δp: 0.342% | |
| Median Δp: -0.591% | |
| 25.0% Δp: -6.307% | |
| 10.0% Δp: -16.931% | |
| 5.0% Δp: -24.574% | |
| 1.0% Δp: -41.728% | |
| 0.1% Δp: -69.844% | |
| Minimum Δp: -98.781% | |
| RMS Δp : 11.741 ± 0.045 % | |
| Same top p: 78.308 ± 0.109 % | |