| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.564700 ± 0.048534 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.96% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.004050 ± 0.000188 | |
| Mean PPL(Q)/PPL(base) : 1.004058 ± 0.000189 | |
| Mean PPL(Q)-PPL(base) : 0.030576 ± 0.001453 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.002095 ± 0.000015 | |
| Maximum KLD: 1.330637 | |
| 99.9% KLD: 0.053552 | |
| 99.0% KLD: 0.014807 | |
| 99.0% KLD: 0.014807 | |
| Median KLD: 0.001325 | |
| 10.0% KLD: 0.000069 | |
| 5.0% KLD: 0.000017 | |
| 1.0% KLD: -0.000001 | |
| Minimum KLD: -0.000220 | |
| ====== Token probability statistics ====== | |
| Mean Δp: 0.024 ± 0.003 % | |
| Maximum Δp: 36.481% | |
| 99.9% Δp: 7.433% | |
| 99.0% Δp: 3.550% | |
| 95.0% Δp: 1.840% | |
| 90.0% Δp: 1.165% | |
| 75.0% Δp: 0.306% | |
| Median Δp: 0.001% | |
| 25.0% Δp: -0.225% | |
| 10.0% Δp: -1.031% | |
| 5.0% Δp: -1.774% | |
| 1.0% Δp: -3.760% | |
| 0.1% Δp: -9.008% | |
| Minimum Δp: -41.338% | |
| RMS Δp : 1.262 ± 0.011 % | |
| Same top p: 97.635 ± 0.040 % | |