llama_model_loader: loaded meta data with 54 key-value pairs and 658 tensors from gemma-4-26B-A4B-it-WIP/gemma-4-26B-A4B-it-Q5_K.gguf (version GGUF V3 (latest))
llama_model_loader: - type  f32:  392 tensors
llama_model_loader: - type q5_1:   60 tensors
llama_model_loader: - type q8_0:    2 tensors
llama_model_loader: - type q4_K:   41 tensors
llama_model_loader: - type q5_K:  154 tensors
llama_model_loader: - type iq4_xs:    9 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q5_K - Medium
print_info: file size   = 16.16 GiB (5.50 BPW) 

====== Perplexity statistics ======
Mean PPL(Q)                   : 16520.065420 ± 329.228992
Mean PPL(base)                : 6617.791728 ± 107.205431
Cor(ln(PPL(Q)), ln(PPL(base))):  90.37%
Mean ln(PPL(Q)/PPL(base))     :   0.914814 ±   0.008723
Mean PPL(Q)/PPL(base)         :   2.496311 ±   0.021774
Mean PPL(Q)-PPL(base)         : 9902.273692 ± 236.836801

====== KL divergence statistics ======
Mean    KLD:   1.337499 ±   0.007116
Maximum KLD:  39.146858
99.9%   KLD:  22.333305
99.0%   KLD:  13.636544
95.0%   KLD:   6.857663
90.0%   KLD:   4.022427
Median  KLD:   0.247420
10.0%   KLD:   0.000893
 5.0%   KLD:   0.000098
 1.0%   KLD:   0.000002
 0.1%   KLD:  -0.000000
Minimum KLD:  -0.000005

====== Token probability statistics ======
Mean    Δp:  0.012 ± 0.038 %
Maximum Δp: 100.000%
99.9%   Δp: 99.650%
99.0%   Δp: 61.136%
95.0%   Δp: 10.497%
90.0%   Δp:  1.636%
75.0%   Δp:  0.005%
Median  Δp: -0.000%
25.0%   Δp: -0.005%
10.0%   Δp: -1.745%
 5.0%   Δp: -10.769%
 1.0%   Δp: -59.710%
 0.1%   Δp: -99.696%
Minimum Δp: -100.000%
RMS Δp    : 14.694 ± 0.097 %
Same top p: 65.841 ± 0.124 %