| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 8.253598 ± 0.051864 | |
| Mean PPL(base) : 7.237090 ± 0.045539 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 97.71% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.131430 ± 0.001346 | |
| Mean PPL(Q)/PPL(base) : 1.140458 ± 0.001535 | |
| Mean PPL(Q)-PPL(base) : 1.016508 ± 0.012175 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.117565 ± 0.000433 | |
| Maximum KLD: 7.079286 | |
| 99.9% KLD: 1.966468 | |
| 99.0% KLD: 0.726076 | |
| 99.0% KLD: 0.726076 | |
| Median KLD: 0.084699 | |
| 10.0% KLD: 0.006988 | |
| 5.0% KLD: 0.002383 | |
| 1.0% KLD: 0.000330 | |
| Minimum KLD: -0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -3.685 ± 0.026 % | |
| Maximum Δp: 69.513% | |
| 99.9% Δp: 34.570% | |
| 99.0% Δp: 17.585% | |
| 95.0% Δp: 7.273% | |
| 90.0% Δp: 3.369% | |
| 75.0% Δp: 0.113% | |
| Median Δp: -0.833% | |
| 25.0% Δp: -6.212% | |
| 10.0% Δp: -15.079% | |
| 5.0% Δp: -21.666% | |
| 1.0% Δp: -38.754% | |
| 0.1% Δp: -69.188% | |
| Minimum Δp: -97.122% | |
| RMS Δp : 10.385 ± 0.045 % | |
| Same top p: 82.770 ± 0.100 % | |