| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 8.029877 ± 0.051646 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 98.94% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.063727 ± 0.000933 | |
| Mean PPL(Q)/PPL(base) : 1.065801 ± 0.000994 | |
| Mean PPL(Q)-PPL(base) : 0.495753 ± 0.008026 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.051126 ± 0.000296 | |
| Maximum KLD: 6.852483 | |
| 99.9% KLD: 1.491291 | |
| 99.0% KLD: 0.440246 | |
| 99.0% KLD: 0.440246 | |
| Median KLD: 0.028683 | |
| 10.0% KLD: 0.002148 | |
| 5.0% KLD: 0.000733 | |
| 1.0% KLD: 0.000110 | |
| Minimum KLD: -0.000059 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -1.501 ± 0.017 % | |
| Maximum Δp: 70.128% | |
| 99.9% Δp: 25.463% | |
| 99.0% Δp: 12.370% | |
| 95.0% Δp: 5.593% | |
| 90.0% Δp: 2.967% | |
| 75.0% Δp: 0.321% | |
| Median Δp: -0.188% | |
| 25.0% Δp: -2.480% | |
| 10.0% Δp: -7.159% | |
| 5.0% Δp: -11.373% | |
| 1.0% Δp: -25.561% | |
| 0.1% Δp: -59.737% | |
| Minimum Δp: -96.409% | |
| RMS Δp : 6.613 ± 0.047 % | |
| Same top p: 89.569 ± 0.081 % | |