| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 17.465106 ± 0.147682 | |
| Mean PPL(base) : 14.009216 ± 0.118474 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 93.39% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.220490 ± 0.003075 | |
| Mean PPL(Q)/PPL(base) : 1.246687 ± 0.003834 | |
| Mean PPL(Q)-PPL(base) : 3.455890 ± 0.056276 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.441957 ± 0.001645 | |
| Maximum KLD: 16.331650 | |
| 99.9% KLD: 5.864139 | |
| 99.0% KLD: 3.003644 | |
| 99.0% KLD: 3.003644 | |
| Median KLD: 0.258535 | |
| 10.0% KLD: 0.008939 | |
| 5.0% KLD: 0.002053 | |
| 1.0% KLD: 0.000174 | |
| Minimum KLD: -0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -4.890 ± 0.044 % | |
| Maximum Δp: 99.975% | |
| 99.9% Δp: 70.255% | |
| 99.0% Δp: 37.236% | |
| 95.0% Δp: 15.056% | |
| 90.0% Δp: 6.624% | |
| 75.0% Δp: 0.255% | |
| Median Δp: -0.337% | |
| 25.0% Δp: -7.613% | |
| 10.0% Δp: -24.787% | |
| 5.0% Δp: -37.890% | |
| 1.0% Δp: -66.824% | |
| 0.1% Δp: -92.729% | |
| Minimum Δp: -99.989% | |
| RMS Δp : 17.470 ± 0.066 % | |
| Same top p: 69.969 ± 0.121 % | |