| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 8.037667 ± 0.050228 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 98.76% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.064696 ± 0.001005 | |
| Mean PPL(Q)/PPL(base) : 1.066835 ± 0.001072 | |
| Mean PPL(Q)-PPL(base) : 0.503543 ± 0.007995 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.066422 ± 0.000322 | |
| Maximum KLD: 6.343029 | |
| 99.9% KLD: 1.563787 | |
| 99.0% KLD: 0.518421 | |
| 99.0% KLD: 0.518421 | |
| Median KLD: 0.041650 | |
| 10.0% KLD: 0.004530 | |
| 5.0% KLD: 0.001681 | |
| 1.0% KLD: 0.000255 | |
| Minimum KLD: 0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.699 ± 0.019 % | |
| Maximum Δp: 69.336% | |
| 99.9% Δp: 26.823% | |
| 99.0% Δp: 11.876% | |
| 95.0% Δp: 4.305% | |
| 90.0% Δp: 1.889% | |
| 75.0% Δp: 0.060% | |
| Median Δp: -0.671% | |
| 25.0% Δp: -4.411% | |
| 10.0% Δp: -10.028% | |
| 5.0% Δp: -14.549% | |
| 1.0% Δp: -28.523% | |
| 0.1% Δp: -62.974% | |
| Minimum Δp: -96.416% | |
| RMS Δp : 7.606 ± 0.046 % | |
| Same top p: 88.690 ± 0.084 % | |