| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 28.239199 ±0.276339 | |
| Mean PPL(base) : 24.931431 ±0.241228 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 97.23% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.124582 ±0.002294 | |
| Mean PPL(Q)/PPL(base) : 1.132675 ±0.002598 | |
| Mean PPL(Q)-PPL(base) : 3.307769 ±0.070198 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.230055 ±0.000737 | |
| Maximum KLD: 8.185512 | |
| 99.9% KLD: 2.553982 | |
| 99.0% KLD: 1.288128 | |
| 99.0% KLD: 1.288128 | |
| Median KLD: 0.147510 | |
| 10.0% KLD: 0.004156 | |
| 5.0% KLD: 0.000713 | |
| 1.0% KLD: 0.000035 | |
| Minimum KLD: -0.000052 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -2.074 ± 0.029 % | |
| Maximum Δp: 95.238% | |
| 99.9% Δp: 55.714% | |
| 99.0% Δp: 30.183% | |
| 95.0% Δp: 12.353% | |
| 90.0% Δp: 5.671% | |
| 75.0% Δp: 0.292% | |
| Median Δp: -0.036% | |
| 25.0% Δp: -3.215% | |
| 10.0% Δp: -13.752% | |
| 5.0% Δp: -22.530% | |
| 1.0% Δp: -43.213% | |
| 0.1% Δp: -69.338% | |
| Minimum Δp: -99.414% | |
| RMS Δp : 11.536 ± 0.050 % | |
| Same top p: 76.903 ± 0.109 % | |