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-Q2_K.gguf (version GGUF V3 (latest))
llama_model_loader: - type  f32:  392 tensors
llama_model_loader: - type q2_K:    1 tensors
llama_model_loader: - type iq1_s:  205 tensors
llama_model_loader: - type iq4_nl:   60 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ2_S - 2.5 bpw
print_info: file size   = 7.39 GiB (2.51 BPW) 

====== Perplexity statistics ======
Mean PPL(Q)                   : 17569.623360 ± 285.143912
Mean PPL(base)                : 6617.791728 ± 107.205431
Cor(ln(PPL(Q)), ln(PPL(base))):  40.59%
Mean ln(PPL(Q)/PPL(base))     :   0.976410 ±   0.017675
Mean PPL(Q)/PPL(base)         :   2.654907 ±   0.046924
Mean PPL(Q)-PPL(base)         : 10951.831633 ± 260.737576

====== KL divergence statistics ======
Mean    KLD:   9.423727 ±   0.015560
Maximum KLD:  54.185097
99.9%   KLD:  39.454365
99.0%   KLD:  28.032539
95.0%   KLD:  20.146471
90.0%   KLD:  17.007847
Median  KLD:   8.712513
10.0%   KLD:   2.362860
 5.0%   KLD:   1.086572
 1.0%   KLD:   0.133040
 0.1%   KLD:   0.006490
Minimum KLD:   0.000003

====== Token probability statistics ======
Mean    Δp: -13.762 ± 0.094 %
Maximum Δp: 99.973%
99.9%   Δp: 96.656%
99.0%   Δp: 69.966%
95.0%   Δp: 13.337%
90.0%   Δp:  2.127%
75.0%   Δp:  0.022%
Median  Δp: -0.000%
25.0%   Δp: -4.540%
10.0%   Δp: -90.575%
 5.0%   Δp: -99.324%
 1.0%   Δp: -99.994%
 0.1%   Δp: -100.000%
Minimum Δp: -100.000%
RMS Δp    : 38.653 ± 0.106 %
Same top p: 10.100 ± 0.079 %