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-Q8_0.gguf (version GGUF V3 (latest))
llama_model_loader: - type  f32:  392 tensors
llama_model_loader: - type  f16:   14 tensors
llama_model_loader: - type q8_0:  251 tensors
llama_model_loader: - type q6_K:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q8_0
print_info: file size   = 24.97 GiB (8.50 BPW) 

====== Perplexity statistics ======
Mean PPL(Q)                   : 17152.384071 ± 340.978253
Mean PPL(base)                : 6617.791728 ± 107.205431
Cor(ln(PPL(Q)), ln(PPL(base))):  93.69%
Mean ln(PPL(Q)/PPL(base))     :   0.952375 ±   0.007362
Mean PPL(Q)/PPL(base)         :   2.591859 ±   0.019081
Mean PPL(Q)-PPL(base)         : 10534.592343 ± 243.443207

====== KL divergence statistics ======
Mean    KLD:   0.580789 ±   0.004459
Maximum KLD:  35.392750
99.9%   KLD:  17.521753
99.0%   KLD:   8.883410
95.0%   KLD:   3.105905
90.0%   KLD:   1.428181
Median  KLD:   0.042224
10.0%   KLD:   0.000139
 5.0%   KLD:   0.000014
 1.0%   KLD:   0.000000
 0.1%   KLD:  -0.000001
Minimum KLD:  -0.000734

====== Token probability statistics ======
Mean    Δp: -0.148 ± 0.025 %
Maximum Δp: 100.000%
99.9%   Δp: 93.766%
99.0%   Δp: 28.845%
95.0%   Δp:  3.859%
90.0%   Δp:  0.508%
75.0%   Δp:  0.001%
Median  Δp: -0.000%
25.0%   Δp: -0.003%
10.0%   Δp: -0.864%
 5.0%   Δp: -4.946%
 1.0%   Δp: -32.363%
 0.1%   Δp: -96.767%
Minimum Δp: -100.000%
RMS Δp    :  9.757 ± 0.093 %
Same top p: 78.799 ± 0.107 %