| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 6.982632 ± 0.049472 | |
| Mean PPL(base) : 6.180577 ± 0.041038 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 97.79% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.122014 ± 0.001510 | |
| Mean PPL(Q)/PPL(base) : 1.129770 ± 0.001706 | |
| Mean PPL(Q)-PPL(base) : 0.802055 ± 0.012688 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.103385 ± 0.000562 | |
| Maximum KLD: 8.783179 | |
| 99.9% KLD: 2.570036 | |
| 99.0% KLD: 0.970630 | |
| 99.0% KLD: 0.970630 | |
| Median KLD: 0.042882 | |
| 10.0% KLD: 0.000696 | |
| 5.0% KLD: 0.000212 | |
| 1.0% KLD: 0.000037 | |
| Minimum KLD: -0.000387 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -0.118 ± 0.024 % | |
| Maximum Δp: 98.177% | |
| 99.9% Δp: 51.694% | |
| 99.0% Δp: 26.477% | |
| 95.0% Δp: 12.678% | |
| 90.0% Δp: 7.513% | |
| 75.0% Δp: 1.666% | |
| Median Δp: 0.011% | |
| 25.0% Δp: -1.220% | |
| 10.0% Δp: -7.227% | |
| 5.0% Δp: -13.450% | |
| 1.0% Δp: -33.309% | |
| 0.1% Δp: -72.790% | |
| Minimum Δp: -98.642% | |
| RMS Δp : 9.341 ± 0.053 % | |
| Same top p: 87.173 ± 0.086 % | |