| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 8.238324 ± 0.051597 | |
| Mean PPL(base) : 6.554978 ± 0.040159 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 95.64% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.228572 ± 0.001835 | |
| Mean PPL(Q)/PPL(base) : 1.256804 ± 0.002306 | |
| Mean PPL(Q)-PPL(base) : 1.683345 ± 0.017653 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.223382 ± 0.000737 | |
| Maximum KLD: 13.166234 | |
| 99.9% KLD: 3.156968 | |
| 99.0% KLD: 1.269592 | |
| 99.0% KLD: 1.269592 | |
| Median KLD: 0.172361 | |
| 10.0% KLD: 0.010869 | |
| 5.0% KLD: 0.003011 | |
| 1.0% KLD: 0.000389 | |
| Minimum KLD: 0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -5.360 ± 0.036 % | |
| Maximum Δp: 79.574% | |
| 99.9% Δp: 44.925% | |
| 99.0% Δp: 25.301% | |
| 95.0% Δp: 11.514% | |
| 90.0% Δp: 5.467% | |
| 75.0% Δp: 0.109% | |
| Median Δp: -1.281% | |
| 25.0% Δp: -9.243% | |
| 10.0% Δp: -22.453% | |
| 5.0% Δp: -32.053% | |
| 1.0% Δp: -54.484% | |
| 0.1% Δp: -87.388% | |
| Minimum Δp: -99.680% | |
| RMS Δp : 14.820 ± 0.056 % | |
| Same top p: 75.373 ± 0.114 % | |