| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.730788 ± 0.049731 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.61% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.025768 ± 0.000567 | |
| Mean PPL(Q)/PPL(base) : 1.026103 ± 0.000582 | |
| Mean PPL(Q)-PPL(base) : 0.196664 ± 0.004583 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.019576 ± 0.000122 | |
| Maximum KLD: 7.128579 | |
| 99.9% KLD: 0.530688 | |
| 99.0% KLD: 0.163194 | |
| 99.0% KLD: 0.163194 | |
| Median KLD: 0.011057 | |
| 10.0% KLD: 0.000745 | |
| 5.0% KLD: 0.000248 | |
| 1.0% KLD: 0.000038 | |
| Minimum KLD: -0.000096 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -0.494 ± 0.010 % | |
| Maximum Δp: 60.888% | |
| 99.9% Δp: 19.533% | |
| 99.0% Δp: 8.890% | |
| 95.0% Δp: 4.150% | |
| 90.0% Δp: 2.365% | |
| 75.0% Δp: 0.435% | |
| Median Δp: -0.028% | |
| 25.0% Δp: -1.070% | |
| 10.0% Δp: -3.708% | |
| 5.0% Δp: -6.092% | |
| 1.0% Δp: -14.010% | |
| 0.1% Δp: -33.372% | |
| Minimum Δp: -79.450% | |
| RMS Δp : 3.873 ± 0.030 % | |
| Same top p: 93.423 ± 0.065 % | |