DeepSeek-R1-Distill-Qwen-7B-GGUF / scores /deepseek-r1-distill-qwen-7b-iq4_nl.log
eaddario's picture
Generate perplexity and kld scores
75442e7 unverified
====== Perplexity statistics ======
Mean PPL(Q) : 25.645935 ± 0.251542
Mean PPL(base) : 24.931431 ± 0.241228
Cor(ln(PPL(Q)), ln(PPL(base))): 99.48%
Mean ln(PPL(Q)/PPL(base)) : 0.028256 ± 0.001007
Mean PPL(Q)/PPL(base) : 1.028659 ± 0.001035
Mean PPL(Q)-PPL(base) : 0.714505 ± 0.027258
====== KL divergence statistics ======
Mean KLD: 0.035420 ± 0.000140
Maximum KLD: 3.406450
99.9% KLD: 0.527580
99.0% KLD: 0.238919
99.0% KLD: 0.238919
Median KLD: 0.019661
10.0% KLD: 0.000409
5.0% KLD: 0.000068
1.0% KLD: 0.000002
Minimum KLD: -0.000216
====== Token probability statistics ======
Mean Δp: -0.143 ± 0.012 %
Maximum Δp: 79.304%
99.9% Δp: 28.190%
99.0% Δp: 14.472%
95.0% Δp: 6.173%
90.0% Δp: 3.204%
75.0% Δp: 0.378%
Median Δp: -0.000%
25.0% Δp: -0.552%
10.0% Δp: -3.690%
5.0% Δp: -6.890%
1.0% Δp: -15.483%
0.1% Δp: -31.224%
Minimum Δp: -74.269%
RMS Δp : 4.584 ± 0.027 %
Same top p: 90.668 ± 0.075 %