Modelpublic

ENOSYS/Kimi-Linear-48B-A3B-Instruct-abliterated-1000-v1-GGUF

sourceHugging Facemitupdated 5mo agoView on Hugging Face

1likes252downloads

Model Card

Experimental global target bits‑per‑weight quantization of huihui-ai/Huihui-Kimi-Linear-48B-A3B-Instruct-abliterated

—Using non-standard (forked) LLaMA C++ branch for quantization.
—Using a CLI tool to build KLD evaluation and imatrix calibration datasets for GGUF models, sourced from eaddario/imatrix-calibration.
—Using dataset sources: tools, math, code, texten, textru.
—Using dataset chunks: 1000.
—Tensors quantinization F16 instead of BF16, Nvidia Pascal architecture friendly like P100.
—Small set of patches added.

Many thanks to Ed Addario for an impressive job.

Quantization comparison

BPW/TGS	SIZE (MiB)	PPL correlation	PPL mean ratio	ΔPPL	Mean KLD	Median KLD	Maximum KLD	99.9% KLD	Mean Δp	RMS Δp
3.000	17574	97.54%	1.099746 ± 0.001198	0.703413 ± 0.008959	0.122940 ± 0.000426	0.066714	9.294135	2.637652	-2.015 ± 0.019 %	10.272 ± 0.039 %
3.250	19038	98.25%	1.064443 ± 0.000969	0.454454 ± 0.006978	0.089130 ± 0.000308	0.050624	10.323897	1.992209	-1.732 ± 0.016 %	8.814 ± 0.035 %
3.500	20502	98.69%	1.042289 ± 0.000821	0.298224 ± 0.005835	0.064519 ± 0.000265	0.032550	8.311751	1.765223	-1.217 ± 0.014 %	7.481 ± 0.033 %
3.750	21966	99.04%	1.030411 ± 0.000697	0.214459 ± 0.004989	0.047054 ± 0.000201	0.022846	8.346386	1.258301	-0.654 ± 0.012 %	6.340 ± 0.031 %
4.000	23430	99.15%	1.028018 ± 0.000653	0.197583 ± 0.004668	0.042084 ± 0.000178	0.020843	7.031458	1.095171	-0.639 ± 0.011 %	5.992 ± 0.029 %
4.250	24894	99.25%	1.021574 ± 0.000610	0.152141 ± 0.004327	0.036486 ± 0.000158	0.018297	6.763760	1.009924	-0.568 ± 0.011 %	5.619 ± 0.028 %
4.500	26358	99.50%	1.012304 ± 0.000495	0.086766 ± 0.003506	0.024419 ± 0.000105	0.012078	5.742187	0.643191	-0.204 ± 0.009 %	4.565 ± 0.023 %
4.750	27822	99.61%	1.007470 ± 0.000434	0.052677 ± 0.003063	0.019040 ± 0.000086	0.009403	4.860535	0.504090	-0.200 ± 0.008 %	4.035 ± 0.020 %
5.000	29286	99.67%	1.006065 ± 0.000397	0.042769 ± 0.002799	0.015935 ± 0.000073	0.008021	3.658825	0.425957	-0.201 ± 0.007 %	3.731 ± 0.020 %