Llama-Guard-3-8B-GGUF / scores /Llama-Guard-3-8B-IQ3_M.log
eaddario's picture
Generate perplexity and kld scores
99b1c6a verified
====== Perplexity statistics ======
Mean PPL(Q) : 7.296410 ± 0.044494
Mean PPL(base) : 6.554978 ± 0.040159
Cor(ln(PPL(Q)), ln(PPL(base))): 97.80%
Mean ln(PPL(Q)/PPL(base)) : 0.107158 ± 0.001284
Mean PPL(Q)/PPL(base) : 1.113110 ± 0.001429
Mean PPL(Q)-PPL(base) : 0.741432 ± 0.009877
====== KL divergence statistics ======
Mean KLD: 0.103618 ± 0.000568
Maximum KLD: 8.069861
99.9% KLD: 2.716604
99.0% KLD: 0.936197
99.0% KLD: 0.936197
Median KLD: 0.055715
10.0% KLD: 0.004556
5.0% KLD: 0.001484
1.0% KLD: 0.000210
Minimum KLD: 0.000001
====== Token probability statistics ======
Mean Δp: -3.332 ± 0.025 %
Maximum Δp: 75.533%
99.9% Δp: 35.065%
99.0% Δp: 14.712%
95.0% Δp: 5.790%
90.0% Δp: 2.603%
75.0% Δp: 0.073%
Median Δp: -0.775%
25.0% Δp: -5.092%
10.0% Δp: -12.092%
5.0% Δp: -18.306%
1.0% Δp: -42.318%
0.1% Δp: -81.231%
Minimum Δp: -99.509%
RMS Δp : 10.006 ± 0.060 %
Same top p: 85.760 ± 0.092 %