| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 6.620408 ± 0.040695 | |
| Mean PPL(base) : 6.554978 ± 0.040159 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.76% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.009932 ± 0.000423 | |
| Mean PPL(Q)/PPL(base) : 1.009982 ± 0.000428 | |
| Mean PPL(Q)-PPL(base) : 0.065430 ± 0.002837 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.011434 ± 0.000081 | |
| Maximum KLD: 2.556788 | |
| 99.9% KLD: 0.366821 | |
| 99.0% KLD: 0.101088 | |
| 99.0% KLD: 0.101088 | |
| Median KLD: 0.006105 | |
| 10.0% KLD: 0.000277 | |
| 5.0% KLD: 0.000078 | |
| 1.0% KLD: 0.000009 | |
| Minimum KLD: -0.000042 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -0.205 ± 0.008 % | |
| Maximum Δp: 59.370% | |
| 99.9% Δp: 18.373% | |
| 99.0% Δp: 8.211% | |
| 95.0% Δp: 3.773% | |
| 90.0% Δp: 2.210% | |
| 75.0% Δp: 0.468% | |
| Median Δp: -0.004% | |
| 25.0% Δp: -0.743% | |
| 10.0% Δp: -2.773% | |
| 5.0% Δp: -4.507% | |
| 1.0% Δp: -9.903% | |
| 0.1% Δp: -27.143% | |
| Minimum Δp: -86.382% | |
| RMS Δp : 3.175 ± 0.030 % | |
| Same top p: 94.641 ± 0.059 % | |