| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 7.539926 ± 0.048262 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.99% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.000770 ± 0.000076 | |
| Mean PPL(Q)/PPL(base) : 1.000770 ± 0.000077 | |
| Mean PPL(Q)-PPL(base) : 0.005802 ± 0.000579 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.000238 ± 0.000002 | |
| Maximum KLD: 0.151414 | |
| 99.9% KLD: 0.005715 | |
| 99.0% KLD: 0.001698 | |
| 99.0% KLD: 0.001698 | |
| Median KLD: 0.000153 | |
| 10.0% KLD: 0.000008 | |
| 5.0% KLD: 0.000001 | |
| 1.0% KLD: -0.000007 | |
| Minimum KLD: -0.000144 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -0.017 ± 0.001 % | |
| Maximum Δp: 14.747% | |
| 99.9% Δp: 2.639% | |
| 99.0% Δp: 1.139% | |
| 95.0% Δp: 0.550% | |
| 90.0% Δp: 0.325% | |
| 75.0% Δp: 0.074% | |
| Median Δp: -0.000% | |
| 25.0% Δp: -0.100% | |
| 10.0% Δp: -0.387% | |
| 5.0% Δp: -0.625% | |
| 1.0% Δp: -1.236% | |
| 0.1% Δp: -2.830% | |
| Minimum Δp: -8.728% | |
| RMS Δp : 0.415 ± 0.004 % | |
| Same top p: 99.218 ± 0.023 % | |