Llama-Guard-3-8B-GGUF / scores /Llama-Guard-3-8B-Q3_K_M.log
eaddario's picture
Generate perplexity and kld scores
5f15980 verified
====== Perplexity statistics ======
Mean PPL(Q) : 7.058128 ± 0.043917
Mean PPL(base) : 6.554978 ± 0.040159
Cor(ln(PPL(Q)), ln(PPL(base))): 98.13%
Mean ln(PPL(Q)/PPL(base)) : 0.073955 ± 0.001199
Mean PPL(Q)/PPL(base) : 1.076758 ± 0.001291
Mean PPL(Q)-PPL(base) : 0.503150 ± 0.008958
====== KL divergence statistics ======
Mean KLD: 0.082098 ± 0.000506
Maximum KLD: 7.242301
99.9% KLD: 2.677090
99.0% KLD: 0.797781
99.0% KLD: 0.797781
Median KLD: 0.039636
10.0% KLD: 0.002107
5.0% KLD: 0.000626
1.0% KLD: 0.000097
Minimum KLD: 0.000000
====== Token probability statistics ======
Mean Δp: -1.456 ± 0.023 %
Maximum Δp: 80.443%
99.9% Δp: 35.332%
99.0% Δp: 17.314%
95.0% Δp: 7.970%
90.0% Δp: 4.563%
75.0% Δp: 0.769%
Median Δp: -0.104%
25.0% Δp: -2.423%
10.0% Δp: -8.113%
5.0% Δp: -13.666%
1.0% Δp: -36.371%
0.1% Δp: -79.787%
Minimum Δp: -99.572%
RMS Δp : 8.722 ± 0.061 %
Same top p: 87.098 ± 0.088 %