| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 15.756085 ± 0.134240 | |
| Mean PPL(base) : 13.052024 ± 0.108483 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 94.45% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.188283 ± 0.002810 | |
| Mean PPL(Q)/PPL(base) : 1.207176 ± 0.003392 | |
| Mean PPL(Q)-PPL(base) : 2.704061 ± 0.047735 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.341147 ± 0.001464 | |
| Maximum KLD: 14.537857 | |
| 99.9% KLD: 5.610807 | |
| 99.0% KLD: 2.699044 | |
| 99.0% KLD: 2.699044 | |
| Median KLD: 0.169474 | |
| 10.0% KLD: 0.004443 | |
| 5.0% KLD: 0.000969 | |
| 1.0% KLD: 0.000097 | |
| Minimum KLD: -0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.851 ± 0.040 % | |
| Maximum Δp: 98.231% | |
| 99.9% Δp: 65.858% | |
| 99.0% Δp: 35.218% | |
| 95.0% Δp: 15.581% | |
| 90.0% Δp: 7.938% | |
| 75.0% Δp: 0.705% | |
| Median Δp: -0.075% | |
| 25.0% Δp: -4.118% | |
| 10.0% Δp: -17.346% | |
| 5.0% Δp: -29.981% | |
| 1.0% Δp: -63.627% | |
| 0.1% Δp: -93.594% | |
| Minimum Δp: -99.962% | |
| RMS Δp : 15.306 ± 0.069 % | |
| Same top p: 75.011 ± 0.114 % | |