| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.301537 ± 0.044767 | |
| Mean PPL(base) : 6.554978 ± 0.040159 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 97.66% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.107860 ± 0.001326 | |
| Mean PPL(Q)/PPL(base) : 1.113892 ± 0.001477 | |
| Mean PPL(Q)-PPL(base) : 0.746559 ± 0.010263 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.108870 ± 0.000594 | |
| Maximum KLD: 7.075473 | |
| 99.9% KLD: 2.822138 | |
| 99.0% KLD: 0.979758 | |
| 99.0% KLD: 0.979758 | |
| Median KLD: 0.057728 | |
| 10.0% KLD: 0.004307 | |
| 5.0% KLD: 0.001381 | |
| 1.0% KLD: 0.000211 | |
| Minimum KLD: 0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -3.144 ± 0.025 % | |
| Maximum Δp: 80.109% | |
| 99.9% Δp: 34.323% | |
| 99.0% Δp: 15.825% | |
| 95.0% Δp: 6.439% | |
| 90.0% Δp: 2.991% | |
| 75.0% Δp: 0.136% | |
| Median Δp: -0.643% | |
| 25.0% Δp: -4.764% | |
| 10.0% Δp: -11.960% | |
| 5.0% Δp: -18.526% | |
| 1.0% Δp: -42.905% | |
| 0.1% Δp: -83.187% | |
| Minimum Δp: -99.496% | |
| RMS Δp : 10.140 ± 0.061 % | |
| Same top p: 85.355 ± 0.093 % | |