llama_model_loader: loaded meta data with 49 key-value pairs and 720 tensors from gemma-4-E4B-it-WIP/gemma-4-E4B-it-Q3_K.gguf (version GGUF V3 (latest))
llama_model_loader: - type  f32:  339 tensors
llama_model_loader: - type  f16:    1 tensors
llama_model_loader: - type q4_1:   11 tensors
llama_model_loader: - type q5_1:    6 tensors
llama_model_loader: - type q8_0:    1 tensors
llama_model_loader: - type q2_K:   11 tensors
llama_model_loader: - type q3_K:  118 tensors
llama_model_loader: - type q4_K:   41 tensors
llama_model_loader: - type q5_K:    6 tensors
llama_model_loader: - type iq3_xxs:   41 tensors
llama_model_loader: - type iq3_s:   49 tensors
llama_model_loader: - type iq4_xs:   96 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 3.06 GiB (3.50 BPW) 

====== Perplexity statistics ======
Mean PPL(Q)                   :  53.460778 ±   0.637355
Mean PPL(base)                :  52.693235 ±   0.610088
Cor(ln(PPL(Q)), ln(PPL(base))):  93.43%
Mean ln(PPL(Q)/PPL(base))     :   0.014461 ±   0.004273
Mean PPL(Q)/PPL(base)         :   1.014566 ±   0.004336
Mean PPL(Q)-PPL(base)         :   0.767544 ±   0.227712

====== KL divergence statistics ======
Mean    KLD:   0.368276 ±   0.001632
Maximum KLD:  16.233898
99.9%   KLD:   6.639840
99.0%   KLD:   2.925996
95.0%   KLD:   1.350669
90.0%   KLD:   0.914818
Median  KLD:   0.166417
10.0%   KLD:   0.001494
 5.0%   KLD:   0.000202
 1.0%   KLD:   0.000008
 0.1%   KLD:   0.000000
Minimum KLD:  -0.000030

====== Token probability statistics ======
Mean    Δp: -1.997 ± 0.038 %
Maximum Δp: 99.669%
99.9%   Δp: 75.373%
99.0%   Δp: 40.206%
95.0%   Δp: 14.981%
90.0%   Δp:  6.041%
75.0%   Δp:  0.341%
Median  Δp: -0.002%
25.0%   Δp: -1.599%
10.0%   Δp: -13.659%
 5.0%   Δp: -27.491%
 1.0%   Δp: -61.928%
 0.1%   Δp: -94.101%
Minimum Δp: -99.998%
RMS Δp    : 14.827 ± 0.072 %
Same top p: 75.953 ± 0.111 %