Llama-Guard-3-8B-GGUF / scores /Llama-Guard-3-8B-Q6_K.log
eaddario's picture
Generate perplexity and kld scores
906cce2 verified
====== Perplexity statistics ======
Mean PPL(Q) : 6.570115 ± 0.040324
Mean PPL(base) : 6.554978 ± 0.040159
Cor(ln(PPL(Q)), ln(PPL(base))): 99.90%
Mean ln(PPL(Q)/PPL(base)) : 0.002306 ± 0.000270
Mean PPL(Q)/PPL(base) : 1.002309 ± 0.000271
Mean PPL(Q)-PPL(base) : 0.015136 ± 0.001781
====== KL divergence statistics ======
Mean KLD: 0.004383 ± 0.000032
Maximum KLD: 2.212321
99.9% KLD: 0.132681
99.0% KLD: 0.034646
99.0% KLD: 0.034646
Median KLD: 0.002591
10.0% KLD: 0.000106
5.0% KLD: 0.000027
1.0% KLD: 0.000002
Minimum KLD: -0.000056
====== Token probability statistics ======
Mean Δp: -0.000 ± 0.005 %
Maximum Δp: 49.226%
99.9% Δp: 11.768%
99.0% Δp: 5.638%
95.0% Δp: 2.842%
90.0% Δp: 1.720%
75.0% Δp: 0.416%
Median Δp: 0.000%
25.0% Δp: -0.391%
10.0% Δp: -1.680%
5.0% Δp: -2.766%
1.0% Δp: -5.808%
0.1% Δp: -15.176%
Minimum Δp: -72.891%
RMS Δp : 2.003 ± 0.019 %
Same top p: 96.391 ± 0.049 %