| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 32.110566 ±0.327098 | |
| Mean PPL(base) : 24.931431 ±0.241228 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 96.81% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.253056 ±0.002558 | |
| Mean PPL(Q)/PPL(base) : 1.287955 ±0.003295 | |
| Mean PPL(Q)-PPL(base) : 7.179135 ±0.111368 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.302751 ±0.000829 | |
| Maximum KLD: 5.970398 | |
| 99.9% KLD: 2.277307 | |
| 99.0% KLD: 1.386654 | |
| 99.0% KLD: 1.386654 | |
| Median KLD: 0.216111 | |
| 10.0% KLD: 0.003771 | |
| 5.0% KLD: 0.000612 | |
| 1.0% KLD: 0.000034 | |
| Minimum KLD: -0.000013 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -1.582 ± 0.032 % | |
| Maximum Δp: 88.233% | |
| 99.9% Δp: 56.689% | |
| 99.0% Δp: 34.469% | |
| 95.0% Δp: 16.136% | |
| 90.0% Δp: 8.038% | |
| 75.0% Δp: 0.436% | |
| Median Δp: -0.027% | |
| 25.0% Δp: -2.816% | |
| 10.0% Δp: -13.759% | |
| 5.0% Δp: -23.729% | |
| 1.0% Δp: -45.043% | |
| 0.1% Δp: -69.818% | |
| Minimum Δp: -99.411% | |
| RMS Δp : 12.323 ± 0.049 % | |
| Same top p: 73.717 ± 0.114 % | |