| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 33.173614 ±0.334504 | |
| Mean PPL(base) : 24.931431 ±0.241228 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 96.33% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.285626 ±0.002707 | |
| Mean PPL(Q)/PPL(base) : 1.330594 ±0.003602 | |
| Mean PPL(Q)-PPL(base) : 8.242184 ±0.120934 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.347117 ±0.000948 | |
| Maximum KLD: 6.902395 | |
| 99.9% KLD: 2.647978 | |
| 99.0% KLD: 1.587362 | |
| 99.0% KLD: 1.587362 | |
| Median KLD: 0.247814 | |
| 10.0% KLD: 0.005469 | |
| 5.0% KLD: 0.000960 | |
| 1.0% KLD: 0.000060 | |
| Minimum KLD: -0.000021 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.845 ± 0.034 % | |
| Maximum Δp: 96.093% | |
| 99.9% Δp: 55.811% | |
| 99.0% Δp: 32.844% | |
| 95.0% Δp: 14.133% | |
| 90.0% Δp: 6.325% | |
| 75.0% Δp: 0.157% | |
| Median Δp: -0.095% | |
| 25.0% Δp: -4.286% | |
| 10.0% Δp: -17.155% | |
| 5.0% Δp: -27.930% | |
| 1.0% Δp: -50.400% | |
| 0.1% Δp: -73.758% | |
| Minimum Δp: -98.444% | |
| RMS Δp : 13.391 ± 0.052 % | |
| Same top p: 71.934 ± 0.116 % | |