| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.706824 ± 0.049523 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.70% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.022664 ± 0.000499 | |
| Mean PPL(Q)/PPL(base) : 1.022922 ± 0.000511 | |
| Mean PPL(Q)-PPL(base) : 0.172700 ± 0.004022 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.015407 ± 0.000095 | |
| Maximum KLD: 4.300019 | |
| 99.9% KLD: 0.426162 | |
| 99.0% KLD: 0.128870 | |
| 99.0% KLD: 0.128870 | |
| Median KLD: 0.008645 | |
| 10.0% KLD: 0.000540 | |
| 5.0% KLD: 0.000171 | |
| 1.0% KLD: 0.000021 | |
| Minimum KLD: -0.000211 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -0.401 ± 0.009 % | |
| Maximum Δp: 64.955% | |
| 99.9% Δp: 17.277% | |
| 99.0% Δp: 8.033% | |
| 95.0% Δp: 3.725% | |
| 90.0% Δp: 2.131% | |
| 75.0% Δp: 0.412% | |
| Median Δp: -0.019% | |
| 25.0% Δp: -0.939% | |
| 10.0% Δp: -3.216% | |
| 5.0% Δp: -5.254% | |
| 1.0% Δp: -11.972% | |
| 0.1% Δp: -31.348% | |
| Minimum Δp: -73.628% | |
| RMS Δp : 3.465 ± 0.030 % | |
| Same top p: 94.228 ± 0.061 % | |