| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 15.685298 ± 0.134793 | |
| Mean PPL(base) : 13.052024 ± 0.108483 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 94.79% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.183781 ± 0.002743 | |
| Mean PPL(Q)/PPL(base) : 1.201752 ± 0.003296 | |
| Mean PPL(Q)-PPL(base) : 2.633274 ± 0.047077 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.321187 ± 0.001430 | |
| Maximum KLD: 14.773328 | |
| 99.9% KLD: 5.503292 | |
| 99.0% KLD: 2.638079 | |
| 99.0% KLD: 2.638079 | |
| Median KLD: 0.153902 | |
| 10.0% KLD: 0.003752 | |
| 5.0% KLD: 0.000800 | |
| 1.0% KLD: 0.000074 | |
| Minimum KLD: -0.000003 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.283 ± 0.039 % | |
| Maximum Δp: 98.523% | |
| 99.9% Δp: 67.160% | |
| 99.0% Δp: 35.229% | |
| 95.0% Δp: 16.196% | |
| 90.0% Δp: 8.554% | |
| 75.0% Δp: 0.886% | |
| Median Δp: -0.042% | |
| 25.0% Δp: -3.441% | |
| 10.0% Δp: -15.597% | |
| 5.0% Δp: -27.773% | |
| 1.0% Δp: -61.892% | |
| 0.1% Δp: -92.942% | |
| Minimum Δp: -99.982% | |
| RMS Δp : 14.828 ± 0.069 % | |
| Same top p: 75.915 ± 0.113 % | |