| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 14.675582 ± 0.124114 | |
| Mean PPL(base) : 14.009216 ± 0.118474 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 98.02% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.046470 ± 0.001681 | |
| Mean PPL(Q)/PPL(base) : 1.047566 ± 0.001761 | |
| Mean PPL(Q)-PPL(base) : 0.666366 ± 0.024756 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.126483 ± 0.000727 | |
| Maximum KLD: 12.042090 | |
| 99.9% KLD: 3.347401 | |
| 99.0% KLD: 1.221437 | |
| 99.0% KLD: 1.221437 | |
| Median KLD: 0.057935 | |
| 10.0% KLD: 0.001389 | |
| 5.0% KLD: 0.000300 | |
| 1.0% KLD: 0.000023 | |
| Minimum KLD: -0.000003 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -1.207 ± 0.024 % | |
| Maximum Δp: 96.911% | |
| 99.9% Δp: 57.065% | |
| 99.0% Δp: 24.734% | |
| 95.0% Δp: 10.112% | |
| 90.0% Δp: 5.090% | |
| 75.0% Δp: 0.486% | |
| Median Δp: -0.024% | |
| 25.0% Δp: -2.250% | |
| 10.0% Δp: -9.426% | |
| 5.0% Δp: -15.820% | |
| 1.0% Δp: -33.452% | |
| 0.1% Δp: -65.695% | |
| Minimum Δp: -99.757% | |
| RMS Δp : 9.254 ± 0.052 % | |
| Same top p: 83.926 ± 0.097 % | |