| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 9.061287 ± 0.057476 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 96.76% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.184569 ± 0.001623 | |
| Mean PPL(Q)/PPL(base) : 1.202700 ± 0.001952 | |
| Mean PPL(Q)-PPL(base) : 1.527163 ± 0.016298 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.164457 ± 0.000676 | |
| Maximum KLD: 9.411176 | |
| 99.9% KLD: 2.855778 | |
| 99.0% KLD: 1.234465 | |
| 99.0% KLD: 1.234465 | |
| Median KLD: 0.104662 | |
| 10.0% KLD: 0.009321 | |
| 5.0% KLD: 0.002701 | |
| 1.0% KLD: 0.000294 | |
| Minimum KLD: 0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -5.112 ± 0.033 % | |
| Maximum Δp: 77.844% | |
| 99.9% Δp: 38.199% | |
| 99.0% Δp: 18.699% | |
| 95.0% Δp: 6.729% | |
| 90.0% Δp: 2.723% | |
| 75.0% Δp: 0.066% | |
| Median Δp: -1.070% | |
| 25.0% Δp: -7.642% | |
| 10.0% Δp: -18.119% | |
| 5.0% Δp: -27.580% | |
| 1.0% Δp: -59.212% | |
| 0.1% Δp: -82.003% | |
| Minimum Δp: -99.539% | |
| RMS Δp : 13.476 ± 0.061 % | |
| Same top p: 81.696 ± 0.102 % | |