eaddario/Llama-Guard-3-8B-GGUF

Experimental layer-wise quantization of meta-llama/Llama-Guard-3-8B

Using LLaMA C++ release b5170 for quantization.

Original model: meta-llama/Llama-Guard-3-8B

From the original model creators:

Llama Guard 3 is a Llama-3.1-8B pretrained model, fine-tuned for content safety classification. Similar to previous versions, it can be used to classify content in both LLM inputs (prompt classification) and in LLM responses (response classification). It acts as an LLM – it generates text in its output that indicates whether a given prompt or response is safe or unsafe, and if unsafe, it also lists the content categories violated.

Llama Guard 3 was aligned to safeguard against the MLCommons standardized hazards taxonomy and designed to support Llama 3.1 capabilities. Specifically, it provides content moderation in 8 languages, and was optimized to support safety and security for search and code interpreter tool calls.

PLEASE READ THIS BEFORE USING THESE EXPERIMENTAL VERSIONS!

An area of personal interest is finding ways to optimize the inference performance of LLMs when deployed in resource-constrained environments like commodity hardware, desktops, laptops, mobiles, edge devices, etc. There are many approaches to accomplish this, including architecture simplification and knowledge distillation, but my focus has been primarily on quantization and pruning.

The method used to produce these experimental versions is covered in Squeezing Tensor Bits: the quest for smaller LLMs, but at a high level it involves using custom versions of llama-imatrix and llama-quantize to identify the influential tensors, and quantize the most important layers to higher bit precision and the less important to lower bits. This process was partly inspired by Dumitru's et al Layer-Wise Quantization: A Pragmatic and Effective Method for Quantizing LLMs Beyond Integer Bit-Levels.

There’re two pull requests (imatrix & quantize) to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified versions will be available on GitHub.

For testing and comparison I'd normally use models produced by Unsloth (Daniel and Michael Han do some really advanced level stuff!) and Bartowski (see credits below), but they don't provide GGUF versions of this model, so all tests and comparisons are done against naive quantizations obtained by simply running llama-quantize with no further optimization.

All experimental versions were generated using an appropriate imatrix created from calibration datasets available at eaddario/imatrix-calibration. At its core, an Importance Matrix (imatrix) is a table or, more broadly, a structured representation that scores the relative importance of different features or parameters in a machine learning model. It essentially quantifies the "impact" each feature has on a specific outcome, prediction, or relationship being modeled, and it helps to counterbalance the negative effects of quantization and pruning.

The process to generate these models is roughly as follows:

1. Convert the the original model's tensors to GGUF F16*
2. Estimate the Perplexity score for the F16 model (baseline) using the wikitext-2-raw-v1 dataset, and save the logits
3. Generate an imatrix from selected calibration datasets
4. Determine tensor and layer Importance Score contribution using a modified version of llama-imatrix
5. Select an appropiate quant level for each tensor using a modified version of llama-quantize
6. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
7. Keep versions with the best scores
8. Repeat until all desired quants are created. I find that quantizations below Q3/IQ3 are not fit for my purposes and therefore do not usually generate them, but happy to provide other quants on request.

*BF16 would be preferred, but Apple's GPUs don't support it yet, and therefore any operations are executed in the CPU, making it unacceptably slow. This is expected to change in the near term but until then, if you are using Apple kit avoid using any models tagged BF16

Models

Sizes (in GB)

Model	Naive	Repo	Shrinkage
Llama-Guard-3-8B-IQ3_M	3.78	3.69	2.4%
Llama-Guard-3-8B-IQ3_S	3.68	3.43	6.8%
Llama-Guard-3-8B-IQ4_NL	4.71	4.39	6.8%
Llama-Guard-3-8B-Q3_K_L	4.32	3.76	13.0%
Llama-Guard-3-8B-Q3_K_M	4.02	3.56	11.4%
Llama-Guard-3-8B-Q3_K_S	3.66	3.31	9.6%
Llama-Guard-3-8B-Q4_K_M	4.92	4.41	10.4%
Llama-Guard-3-8B-Q4_K_S	4.69	4.28	8.7%
Llama-Guard-3-8B-Q5_K_M	5.73	5.38	6.1%
Llama-Guard-3-8B-Q5_K_S	5.60	5.24	6.4%
Llama-Guard-3-8B-Q6_K	6.60	6.57	0.5%
Llama-Guard-3-8B-Q8_0	8.54	7.73	9.5%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
Llama-Guard-3-8B-IQ3_M	7.423790 ±0.046691	97.11%	0.134115 ±0.000651	11.077 ±0.059
Llama-Guard-3-8B-IQ3_S	7.746531 ±0.048960	96.22%	0.179586 ±0.000744	12.616 ±0.060
Llama-Guard-3-8B-IQ4_NL	6.935864 ±0.042688	98.71%	0.059280 ±0.000325	7.170 ±0.046
Llama-Guard-3-8B-Q3_K_L	7.630634 ±0.047920	96.28%	0.165526 ±0.000769	12.135 ±0.061
Llama-Guard-3-8B-Q3_K_M	7.831542 ±0.049335	95.77%	0.188482 ±0.000852	12.979 ±0.062
Llama-Guard-3-8B-Q3_K_S	8.269311 ±0.052149	94.63%	0.239987 ±0.001029	14.794 ±0.066
Llama-Guard-3-8B-Q4_K_M	6.908041 ±0.042539	98.78%	0.055843 ±0.000320	7.016 ±0.048
Llama-Guard-3-8B-Q4_K_M (naive)	6.731828 ±0.041532	99.34%	0.030829 ±0.000214	5.255 ±0.045
Llama-Guard-3-8B-Q4_K_S	6.930856 ±0.042651	98.70%	0.059620 ±0.000336	7.285 ±0.049
Llama-Guard-3-8B-Q5_K_M	6.648795 ±0.040800	99.62%	0.017289 ±0.000115	3.870 ±0.034
Llama-Guard-3-8B-Q5_K_S	6.659786 ±0.040894	99.60%	0.018179 ±0.000120	3.957 ±0.034
Llama-Guard-3-8B-Q6_K	6.581335 ±0.040401	99.83%	0.007279 ±0.000061	2.532 ±0.028
Llama-Guard-3-8B-Q8_0	6.569465 ±0.040265	99.89%	0.004781 ±0.000042	2.072 ±0.025
Llama-Guard-3-8B-F16	6.554978 ±0.040159	100%	N/A	N/A

ARC, HellaSwag, MMLU, Truthful QA and WinoGrande scores

Scores generated using llama-perplexity with 750 tasks per test, and a context size of 768 tokens.

For the test data used in the generation of these scores, follow the appropiate links: HellaSwag, ARC, MMLU, Truthful QA and WinoGrande

Model	ARC	HellaSwag	MMLU	Truthful QA	WinoGrande	Avg Score
Llama-Guard-3-8B-IQ3_M	66.5333 ±1.7242	80.40	36.6667 ±1.7608	31.4667 ±1.6968	73.2000 ±1.6184	57.65
Llama-Guard-3-8B-IQ3_S	65.4667 ±1.7374	79.07	35.2000 ±1.7451	29.2000 ±1.6614	70.8000 ±1.6614	55.95
Llama-Guard-3-8B-IQ4_NL	64.9333 ±1.7436	79.60	36.8000 ±1.7621	30.5333 ±1.6828	73.2000 ±1.6184	57.01
Llama-Guard-3-8B-Q3_K_L	64.9333 ±1.7436	78.93	37.0667 ±1.7648	33.8667 ±1.7292	72.2667 ±1.6358	57.41
Llama-Guard-3-8B-Q3_K_M	63.6000 ±1.7581	78.67	36.6667 ±1.7608	33.4667 ±1.7242	70.6667 ±1.6636	56.61
Llama-Guard-3-8B-Q3_K_S	60.2667 ±1.7880	77.46	35.4667 ±1.7481	34.1333 ±1.7325	71.7333 ±1.6453	55.81
Llama-Guard-3-8B-Q4_K_M	65.6000 ±1.7358	80.26	38.1333 ±1.7748	30.4000 ±1.6807	72.2667 ±1.6358	57.33
Llama-Guard-3-8B-Q4_K_M (naive)	66.9786 ±1.7207	79.20	40.2667 ±1.7920	31.0559 ±2.5827	74.2667 ±1.5974	58.35
Llama-Guard-3-8B-Q4_K_S	66.1333 ±1.7292	80.00	37.8667 ±1.7724	30.4000 ±1.6807	71.6000 ±1.6477	57.20
Llama-Guard-3-8B-Q5_K_M	65.8667 ±1.7325	81.33	38.0000 ±1.7736	31.6000 ±1.6988	72.6667 ±1.6284	57.89
Llama-Guard-3-8B-Q5_K_S	65.7333 ±1.7342	81.33	37.4667 ±1.7686	31.8667 ±1.7026	72.9333 ±1.6235	57.87
Llama-Guard-3-8B-Q6_K	65.6000 ±1.7358	81.06	38.6667 ±1.7794	30.9333 ±1.6889	72.5333 ±1.6309	57.76
Llama-Guard-3-8B-Q8_0	65.3333 ±1.7389	81.60	38.4000 ±1.7771	30.8000 ±1.6869	72.8000 ±1.6260	57.79
Llama-Guard-3-8B-F16	64.9333 ±1.7436	81.60	38.2667 ±1.7759	30.6667 ±1.6849	72.8000 ±1.6260	57.65

Tokens per Second - Benchmarks

Scores generated using llama-bench. Naive (llama-quantize with no optimization) Q4_K_M quantization included for comparison.

model	size	params	backend	threads	test	t/s
Llama-Guard-3-8B-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	pp512	312.18 ± 0.88
Llama-Guard-3-8B-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	tg128	27.88 ± 0.03
Llama-Guard-3-8B-Q4_K_M	4.10 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	44.53 ± 0.11
Llama-Guard-3-8B-Q4_K_M (naive)	4.58 GiB	8.03 B	Metal,BLAS	6	pp512	329.30 ± 0.12
Llama-Guard-3-8B-Q4_K_M (naive)	4.58 GiB	8.03 B	Metal,BLAS	6	tg128	26.51 ± 0.02
Llama-Guard-3-8B-Q4_K_M (naive)	4.58 GiB	8.03 B	Metal,BLAS	6	pp1024+tg1024	42.69 ± 1.00

Metrics used

Perplexity: one of the key metrics used in NLP evaluation. It measures the quality of a language model by evaluating how well it predicts the next token given a particular sequence of words. A PPL of 1 indicates an exact match between predicted and actual, whereas values greater than one indicate a degree of "surprise" the generated token differs from the expected.

Kullback–Leibler (KL) Divergence: a statistical measure of how much a probability distribution differs from another. When quantizing models (or altering the original tensors in any way for that matter), the closest we can preserve the weights' probability distribution to the original model the better, thus the closest to 0 the better.

AI2 Reasoning Challenge (ARC): a benchmark to evaluate the ability of AI models to answer complex science questions that require logical reasoning beyond pattern matching.

HellaSwag: the Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations (bit of a mouthful!) is a benchmark designed to test commonsense natural language inference. It requires the model to predict the most likely ending of a sentence.

MMLU: the Massive Multitask Language Understanding evaluates LLMs’ general knowledge and problem-solving abilities across 57 subjects, including elementary mathematics, US history, computer science, and law.

Truthful QA: evaluates how well LLMs generate truthful responses to questions. It identifies whether AI models can avoid generating false or misleading information, particularly in areas where human knowledge is prone to misconceptions.

Winogrande: based on the Winograd Schema Challenge, is a natural language understanding task requiring models to resolve ambiguities in sentences involving pronoun references.

Credits

A big Thank You! to Colin Kealty for the many contributions and for being one of the best sources of high quality quantized models available in Hugginface, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries.