Llama-Guard-3-8B-GGUF / scores /Llama-Guard-3-8B-Q3_K_L.log
eaddario's picture
Generate perplexity and kld scores
aabdfea verified
====== Perplexity statistics ======
Mean PPL(Q) : 6.983531 ± 0.043462
Mean PPL(base) : 6.554978 ± 0.040159
Cor(ln(PPL(Q)), ln(PPL(base))): 98.37%
Mean ln(PPL(Q)/PPL(base)) : 0.063330 ± 0.001119
Mean PPL(Q)/PPL(base) : 1.065378 ± 0.001192
Mean PPL(Q)-PPL(base) : 0.428552 ± 0.008232
====== KL divergence statistics ======
Mean KLD: 0.071002 ± 0.000463
Maximum KLD: 7.889337
99.9% KLD: 2.299609
99.0% KLD: 0.707767
99.0% KLD: 0.707767
Median KLD: 0.033880
10.0% KLD: 0.001714
5.0% KLD: 0.000489
1.0% KLD: 0.000073
Minimum KLD: -0.000059
====== Token probability statistics ======
Mean Δp: -1.167 ± 0.021 %
Maximum Δp: 85.060%
99.9% Δp: 33.978%
99.0% Δp: 16.661%
95.0% Δp: 7.783%
90.0% Δp: 4.533%
75.0% Δp: 0.828%
Median Δp: -0.060%
25.0% Δp: -2.077%
10.0% Δp: -7.263%
5.0% Δp: -12.280%
1.0% Δp: -32.484%
0.1% Δp: -77.734%
Minimum Δp: -98.234%
RMS Δp : 8.106 ± 0.060 %
Same top p: 87.894 ± 0.086 %