| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 8.963688 ± 0.058386 | |
| Mean PPL(base) : 7.237090 ± 0.045539 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 95.93% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.213963 ± 0.001840 | |
| Mean PPL(Q)/PPL(base) : 1.238576 ± 0.002279 | |
| Mean PPL(Q)-PPL(base) : 1.726598 ± 0.019534 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.209768 ± 0.000734 | |
| Maximum KLD: 11.045219 | |
| 99.9% KLD: 3.037534 | |
| 99.0% KLD: 1.327696 | |
| 99.0% KLD: 1.327696 | |
| Median KLD: 0.145406 | |
| 10.0% KLD: 0.013515 | |
| 5.0% KLD: 0.004453 | |
| 1.0% KLD: 0.000589 | |
| Minimum KLD: 0.000000 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -4.187 ± 0.035 % | |
| Maximum Δp: 87.898% | |
| 99.9% Δp: 53.785% | |
| 99.0% Δp: 29.326% | |
| 95.0% Δp: 12.210% | |
| 90.0% Δp: 5.784% | |
| 75.0% Δp: 0.302% | |
| Median Δp: -0.896% | |
| 25.0% Δp: -7.479% | |
| 10.0% Δp: -19.150% | |
| 5.0% Δp: -28.492% | |
| 1.0% Δp: -52.697% | |
| 0.1% Δp: -83.945% | |
| Minimum Δp: -97.765% | |
| RMS Δp : 13.969 ± 0.056 % | |
| Same top p: 77.664 ± 0.110 % | |