DeepSeek-R1-Distill-Qwen-7B-GGUF / scores /deepseek-r1-distill-qwen-7b-iq3_s.log
eaddario's picture
Generate perplexity and kld scores
75442e7 unverified
raw
history blame
1.07 kB
====== Perplexity statistics ======
Mean PPL(Q) : 26.638550 ± 0.259075
Mean PPL(base) : 24.931431 ± 0.241228
Cor(ln(PPL(Q)), ln(PPL(base))): 98.51%
Mean ln(PPL(Q)/PPL(base)) : 0.066230 ± 0.001677
Mean PPL(Q)/PPL(base) : 1.068473 ± 0.001792
Mean PPL(Q)-PPL(base) : 1.707120 ± 0.046752
====== KL divergence statistics ======
Mean KLD: 0.115060 ± 0.000427
Maximum KLD: 6.862354
99.9% KLD: 1.668972
99.0% KLD: 0.747263
99.0% KLD: 0.747263
Median KLD: 0.065809
10.0% KLD: 0.001661
5.0% KLD: 0.000301
1.0% KLD: 0.000018
Minimum KLD: -0.000103
====== Token probability statistics ======
Mean Δp: -1.254 ± 0.021 %
Maximum Δp: 92.550%
99.9% Δp: 43.219%
99.0% Δp: 21.697%
95.0% Δp: 8.707%
90.0% Δp: 4.022%
75.0% Δp: 0.260%
Median Δp: -0.015%
25.0% Δp: -1.839%
10.0% Δp: -8.696%
5.0% Δp: -15.122%
1.0% Δp: -31.087%
0.1% Δp: -57.206%
Minimum Δp: -96.125%
RMS Δp : 8.252 ± 0.042 %
Same top p: 83.845 ± 0.095 %