DeepSeek-R1-Distill-Qwen-7B-GGUF / scores /deepseek-r1-distill-qwen-7b-q5_k_m.log
eaddario's picture
Generate perplexity and kld scores
75442e7 unverified
====== Perplexity statistics ======
Mean PPL(Q) : 25.331908 ± 0.247609
Mean PPL(base) : 24.931431 ± 0.241228
Cor(ln(PPL(Q)), ln(PPL(base))): 99.81%
Mean ln(PPL(Q)/PPL(base)) : 0.015935 ± 0.000613
Mean PPL(Q)/PPL(base) : 1.016063 ± 0.000623
Mean PPL(Q)-PPL(base) : 0.400477 ± 0.016493
====== KL divergence statistics ======
Mean KLD: 0.008227 ± 0.000037
Maximum KLD: 3.104945
99.9% KLD: 0.117209
99.0% KLD: 0.052437
99.0% KLD: 0.052437
Median KLD: 0.004830
10.0% KLD: 0.000092
5.0% KLD: 0.000012
1.0% KLD: -0.000002
Minimum KLD: -0.000180
====== Token probability statistics ======
Mean Δp: -0.032 ± 0.006 %
Maximum Δp: 38.164%
99.9% Δp: 13.916%
99.0% Δp: 6.928%
95.0% Δp: 3.133%
90.0% Δp: 1.652%
75.0% Δp: 0.209%
Median Δp: -0.000%
25.0% Δp: -0.248%
10.0% Δp: -1.741%
5.0% Δp: -3.275%
1.0% Δp: -7.358%
0.1% Δp: -14.696%
Minimum Δp: -60.848%
RMS Δp : 2.211 ± 0.014 %
Same top p: 95.381 ± 0.054 %