Llama-Guard-3-8B-GGUF / scores /Llama-Guard-3-8B-IQ4_NL.log
eaddario's picture
Generate perplexity and kld scores
402e63c verified
====== Perplexity statistics ======
Mean PPL(Q) : 6.763095 ± 0.041790
Mean PPL(base) : 6.554978 ± 0.040159
Cor(ln(PPL(Q)), ln(PPL(base))): 99.26%
Mean ln(PPL(Q)/PPL(base)) : 0.031256 ± 0.000750
Mean PPL(Q)/PPL(base) : 1.031749 ± 0.000774
Mean PPL(Q)-PPL(base) : 0.208116 ± 0.005240
====== KL divergence statistics ======
Mean KLD: 0.034552 ± 0.000246
Maximum KLD: 4.819600
99.9% KLD: 1.188445
99.0% KLD: 0.322914
99.0% KLD: 0.322914
Median KLD: 0.017059
10.0% KLD: 0.000839
5.0% KLD: 0.000242
1.0% KLD: 0.000033
Minimum KLD: -0.000013
====== Token probability statistics ======
Mean Δp: -0.587 ± 0.015 %
Maximum Δp: 75.339%
99.9% Δp: 26.308%
99.0% Δp: 12.292%
95.0% Δp: 5.874%
90.0% Δp: 3.437%
75.0% Δp: 0.690%
Median Δp: -0.025%
25.0% Δp: -1.314%
10.0% Δp: -4.700%
5.0% Δp: -7.828%
1.0% Δp: -20.635%
0.1% Δp: -55.049%
Minimum Δp: -97.416%
RMS Δp : 5.614 ± 0.049 %
Same top p: 91.354 ± 0.074 %