| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 9.156208 ± 0.059074 | |
| Mean PPL(base) : 7.534124 ± 0.048206 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 96.88% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.194990 ± 0.001607 | |
| Mean PPL(Q)/PPL(base) : 1.215298 ± 0.001953 | |
| Mean PPL(Q)-PPL(base) : 1.622084 ± 0.017205 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.176807 ± 0.000516 | |
| Maximum KLD: 7.850962 | |
| 99.9% KLD: 2.159336 | |
| 99.0% KLD: 0.866296 | |
| 99.0% KLD: 0.866296 | |
| Median KLD: 0.144583 | |
| 10.0% KLD: 0.010968 | |
| 5.0% KLD: 0.003578 | |
| 1.0% KLD: 0.000530 | |
| Minimum KLD: 0.000004 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -4.620 ± 0.031 % | |
| Maximum Δp: 81.902% | |
| 99.9% Δp: 38.750% | |
| 99.0% Δp: 22.129% | |
| 95.0% Δp: 9.587% | |
| 90.0% Δp: 4.291% | |
| 75.0% Δp: 0.093% | |
| Median Δp: -1.101% | |
| 25.0% Δp: -8.157% | |
| 10.0% Δp: -19.324% | |
| 5.0% Δp: -27.170% | |
| 1.0% Δp: -44.631% | |
| 0.1% Δp: -74.280% | |
| Minimum Δp: -96.184% | |
| RMS Δp : 12.577 ± 0.047 % | |
| Same top p: 77.814 ± 0.110 % | |