| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 8.869361 ± 0.056188 | |
| Mean PPL(base) : 7.237090 ± 0.045539 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 96.40% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.203384 ± 0.001694 | |
| Mean PPL(Q)/PPL(base) : 1.225543 ± 0.002076 | |
| Mean PPL(Q)-PPL(base) : 1.632272 ± 0.017247 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.171689 ± 0.000675 | |
| Maximum KLD: 8.647476 | |
| 99.9% KLD: 3.093943 | |
| 99.0% KLD: 1.167801 | |
| 99.0% KLD: 1.167801 | |
| Median KLD: 0.116922 | |
| 10.0% KLD: 0.009604 | |
| 5.0% KLD: 0.003321 | |
| 1.0% KLD: 0.000607 | |
| Minimum KLD: 0.000004 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -5.020 ± 0.030 % | |
| Maximum Δp: 68.248% | |
| 99.9% Δp: 34.217% | |
| 99.0% Δp: 17.833% | |
| 95.0% Δp: 7.053% | |
| 90.0% Δp: 2.939% | |
| 75.0% Δp: 0.031% | |
| Median Δp: -1.308% | |
| 25.0% Δp: -7.961% | |
| 10.0% Δp: -18.717% | |
| 5.0% Δp: -26.619% | |
| 1.0% Δp: -49.119% | |
| 0.1% Δp: -82.457% | |
| Minimum Δp: -99.092% | |
| RMS Δp : 12.587 ± 0.055 % | |
| Same top p: 80.614 ± 0.104 % | |