| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 23.315818 ± 0.220552 | |
| Mean PPL(base) : 23.352232 ± 0.220841 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 99.50% | |
| Mean ln(PPL(Q)/PPL(base)) : -0.001561 ± 0.000949 | |
| Mean PPL(Q)/PPL(base) : 0.998441 ± 0.000948 | |
| Mean PPL(Q)-PPL(base) : -0.036415 ± 0.022157 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.037680 ± 0.000153 | |
| Maximum KLD: 5.026946 | |
| 99.9% KLD: 0.544697 | |
| 99.0% KLD: 0.251806 | |
| 99.0% KLD: 0.251806 | |
| Median KLD: 0.018222 | |
| 10.0% KLD: 0.000471 | |
| 5.0% KLD: 0.000107 | |
| 1.0% KLD: -0.000012 | |
| Minimum KLD: -0.000554 | |
| ====== Token probability statistics ====== | |
| Mean Δp: 0.151 ± 0.012 % | |
| Maximum Δp: 80.430% | |
| 99.9% Δp: 29.435% | |
| 99.0% Δp: 16.228% | |
| 95.0% Δp: 7.580% | |
| 90.0% Δp: 4.031% | |
| 75.0% Δp: 0.584% | |
| Median Δp: 0.000% | |
| 25.0% Δp: -0.350% | |
| 10.0% Δp: -3.338% | |
| 5.0% Δp: -6.778% | |
| 1.0% Δp: -15.802% | |
| 0.1% Δp: -30.165% | |
| Minimum Δp: -63.405% | |
| RMS Δp : 4.833 ± 0.025 % | |
| Same top p: 91.410 ± 0.072 % | |
| llama_perf_context_print: load time = 103826.14 ms | |
| llama_perf_context_print: prompt eval time = 1941845.24 ms / 304128 tokens ( 6.38 ms per token, 156.62 tokens per second) | |
| llama_perf_context_print: eval time = 0.00 ms / 1 runs ( 0.00 ms per token, inf tokens per second) | |
| llama_perf_context_print: total time = 2633698.11 ms / 304129 tokens | |