| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.003106 ± 0.043412 | |
| Mean PPL(base) : 6.554978 ± 0.040159 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 98.57% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.066129 ± 0.001044 | |
| Mean PPL(Q)/PPL(base) : 1.068365 ± 0.001115 | |
| Mean PPL(Q)-PPL(base) : 0.448128 ± 0.007768 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.069975 ± 0.000336 | |
| Maximum KLD: 5.985790 | |
| 99.9% KLD: 1.647494 | |
| 99.0% KLD: 0.496019 | |
| 99.0% KLD: 0.496019 | |
| Median KLD: 0.046161 | |
| 10.0% KLD: 0.001926 | |
| 5.0% KLD: 0.000530 | |
| 1.0% KLD: 0.000071 | |
| Minimum KLD: -0.000000 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -1.364 ± 0.020 % | |
| Maximum Δp: 80.174% | |
| 99.9% Δp: 32.730% | |
| 99.0% Δp: 17.965% | |
| 95.0% Δp: 8.831% | |
| 90.0% Δp: 4.993% | |
| 75.0% Δp: 0.697% | |
| Median Δp: -0.125% | |
| 25.0% Δp: -2.890% | |
| 10.0% Δp: -8.969% | |
| 5.0% Δp: -13.891% | |
| 1.0% Δp: -27.946% | |
| 0.1% Δp: -63.303% | |
| Minimum Δp: -97.452% | |
| RMS Δp : 7.869 ± 0.046 % | |
| Same top p: 86.071 ± 0.091 % | |