eaddario/Dolphin3.0-Mistral-24B-GGUF

Experimental GGUF quantized versions of cognitivecomputations/Dolphin3.0-Mistral-24B

Using LLaMA C++ release b4837 for quantization.

Original model: cognitivecomputations/Dolphin3.0-Mistral-24B

From the original model creators:

Dolphin 3.0 is the next generation of the Dolphin series of instruct-tuned models. Designed to be the ultimate general purpose local model, enabling coding, math, agentic, function calling, and general use cases.

Dolphin aims to be a general purpose instruct model, similar to the models behind ChatGPT, Claude, Gemini. But these models present problems for businesses seeking to include AI in their products.

1. They maintain control of the system prompt, deprecating and changing things as they wish, often causing software to break.
2. They maintain control of the model versions, sometimes changing things silently, or deprecating older models that your business relies on.
3. They maintain control of the alignment, and in particular the alignment is one-size-fits all, not tailored to the application.
4. They can see all your queries and they can potentially use that data in ways you wouldn't want. Dolphin, in contrast, is steerable and gives control to the system owner. You set the system prompt. You decide the alignment. You have control of your data. Dolphin does not impose its ethics or guidelines on you. You are the one who decides the guidelines.

Dolphin belongs to YOU, it is your tool, an extension of your will. Just as you are personally responsible for what you do with a knife, gun, fire, car, or the internet, you are the creator and originator of any content you generate with Dolphin.

From Eric Hartford's, the creator of the Dolphin model series, Uncensored Models:

Most of these models (for example, Alpaca, Vicuna, WizardLM, MPT-7B-Chat, Wizard-Vicuna, GPT4-X-Vicuna) have some sort of embedded alignment. For general purposes, this is a good thing. This is what stops the model from doing bad things, like teaching you how to cook meth and make bombs. But what is the nature of this alignment? And, why is it so?

The reason these models are aligned is that they are trained with data that was generated by ChatGPT, which itself is aligned by an alignment team at OpenAI. As it is a black box, we don't know all the reasons for the decisions that were made, but we can observe it generally is aligned with American popular culture, and to obey American law, and with a liberal and progressive political bias.

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, edge devices, etc. There are many approaches to accomplish this, including architecture simplification and knowledge distillation, but for now I'm focusing primarily on quantization and pruning.

The process of quantization reduces the precision of the model's weights, leading to significant reductions in model size, memory needs and computational requirements (a good thing), but this however comes at the expense of a loss in the model's capabilities and accuracy (a bad thing!).

Another approach is to prune the model, that is, to selectively zero-out groups of parameters. Although significant reductions can be achieved this way, the risk of severely degrading the model's performance is markedly higher than when quantizing, as the process requires a deep understanding of the model's architecture in order to identify which tensors can be safely zero'ed. For all means and purposes, pruning is the equivalent of lobotomizing the LLM!

A successful outcome is when the overall size is reduced with no, or negligible, loss of capabilities (i.e. language understanding, math and logic problem-solving, conversation, coding, domain-specific knowledge, etc.) compared to the original version. On that regard, the method I'm using seems to yield some modest but encouraging results, and the versions available in this repo are on average 3% smaller than other, high-quality, sources with negligible loss of capability. As I continue to improve the process and develop tools to automate it, I aim to achieve further reductions in the 10-15% range, maybe more.

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 only the latter offers a version of these model so all tests and comparisons are done against a single reference.

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. Quantize & prune versions of the base model
5. Calculate Perplexity, KL Divergence, ARC (Easy+Challenge), HellaSwag, MMLU, Truthful QA and WinoGrande scores for each quantized model
6. Keep versions with the best scores
7. 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	Bartowski	Repo	Shrinkage
Dolphin3.0-Mistral-24B-IQ3_M	10.65	10.25	3.8%
Dolphin3.0-Mistral-24B-IQ3_S	N/A	10.03	N/A
Dolphin3.0-Mistral-24B-IQ4_NL	13.05	13.05	0%
Dolphin3.0-Mistral-24B-Q3_K_L	12.40	12.00	0%
Dolphin3.0-Mistral-24B-Q3_K_M	11.47	11.08	3.4%
Dolphin3.0-Mistral-24B-Q3_K_S	10.40	10.00	3.8%
Dolphin3.0-Mistral-24B-Q4_K_M	14.33	13.99	2.4%
Dolphin3.0-Mistral-24B-Q4_K_S	13.55	13.13	3.1%
Dolphin3.0-Mistral-24B-Q5_K_M	16.76	16.26	3%
Dolphin3.0-Mistral-24B-Q5_K_S	16.30	15.80	3.1%
Dolphin3.0-Mistral-24B-Q6_K	19.35	18.75	3.1%
Dolphin3.0-Mistral-24B-Q8_0	25.05	24.14	3.6%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
Dolphin3.0-Mistral-24B-IQ3_M	8.536738 ±0.059988	98.03%	0.103999 ±0.000378	9.251 ±0.039
Dolphin3.0-Mistral-24B-IQ3_S	8.602292 ±0.060945	97.95%	0.108536 ±0.000393	9.309 ±0.040
Dolphin3.0-Mistral-24B-IQ4_NL	7.895583 ±0.054759	99.38%	0.031400 ±0.000136	5.278 ±0.027
Dolphin3.0-Mistral-24B-Q3_K_L	8.320959 ±0.058079	98.43%	0.083143 ±0.000296	8.464 ±0.035
Dolphin3.0-Mistral-24B-Q3_K_M	8.389337 ±0.058940	98.30%	0.089559 ±0.000330	8.734 ±0.037
Dolphin3.0-Mistral-24B-Q3_K_S	8.681563 ±0.061366	97.26%	0.138030 ±0.000557	10.731 ±0.047
Dolphin3.0-Mistral-24B-Q4_K_M	7.868503 ±0.054510	99.45%	0.027967 ±0.000110	4.959 ±0.024
Dolphin3.0-Mistral-24B-Q4_K_S	7.922314 ±0.055044	99.40%	0.030745 ±0.000128	5.176 ±0.026
Dolphin3.0-Mistral-24B-Q5_K_M	7.785290 ±0.053616	99.63%	0.018900 ±0.000070	4.199 ±0.019
Dolphin3.0-Mistral-24B-Q5_K_S	7.819818 ±0.053887	99.62%	0.019857 ±0.000076	4.338 ±0.020
Dolphin3.0-Mistral-24B-Q6_K	7.757601 ±0.053375	99.70%	0.015707 ±0.000047	3.852 ±0.014
Dolphin3.0-Mistral-24B-Q8_0	7.737414 ±0.053007	99.72%	0.014644 ±0.000046	3.754 ±0.015
Dolphin3.0-Mistral-24B-F16	9.366577 ±0.066397	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. Naive (llama-quantize with no optimization) Q4_K_M quantization included for comparison.

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
Dolphin3.0-Mistral-24B-IQ3_M	70.5333 ±1.6658	80.80	43.7333 ±1.8126	35.4667 ±1.7481	74.9333 ±1.5836
Dolphin3.0-Mistral-24B-IQ3_S	70.1333 ±1.6723	80.93	43.7333 ±1.8126	36.5333 ±1.7594	74.4000 ±1.5947
Dolphin3.0-Mistral-24B-IQ4_NL	72.1333 ±1.6382	80.27	42.6667 ±1.8072	35.7333 ±1.7510	76.5333 ±1.5485
Dolphin3.0-Mistral-24B-Q3_K_L	72.5333 ±1.6309	80.93	41.6000 ±1.8010	34.2667 ±1.7342	75.7333 ±1.5664
Dolphin3.0-Mistral-24B-Q3_K_M	73.4667 ±1.6132	80.93	42.1333 ±1.8042	34.9300 ±1.6774	76.1333 ±1.5576
Dolphin3.0-Mistral-24B-Q3_K_S	70.5333 ±1.6658	80.67	41.2000 ±1.7984	35.2000 ±1.7451	74.6667 ±1.5892
Dolphin3.0-Mistral-24B-Q4_K_M	72.6667 ±1.6284	81.73	42.5333 ±1.8065	35.0667 ±1.7436	77.6000 ±1.5234
Dolphin3.0-Mistral-24B-Q4_K_M-bartowski	72.2667 ±1.6358	81.73	42.8000 ±1.8079	35.0667 ±1.7436	76.8000 ±1.5424
Dolphin3.0-Mistral-24B-Q4_K_S	72.4000 ±1.6334	81.47	43.0667 ±1.8093	35.3333 ±1.7466	77.2000 ±1.5330
Dolphin3.0-Mistral-24B-Q5_K_M	72.0000 ±1.6406	81.20	42.9333 ±1.8086	35.7333 ±1.7510	78.1333 ±1.5103
Dolphin3.0-Mistral-24B-Q5_K_S	72.6667 ±1.6284	81.47	41.6000 ±1.8010	35.7333 ±1.7510	76.5333 ±1.5485
Dolphin3.0-Mistral-24B-Q6_K	72.4000 ±1.6334	81.47	43.0667 ±1.8093	36.6667 ±1.7608	77.8667 ±1.5169
Dolphin3.0-Mistral-24B-Q8_0	72.8000 ±1.6260	81.33	43.0667 ±1.8093	35.7333 ±1.7510	77.4667 ±1.5266
Dolphin3.0-Mistral-24B-F16	71.6000 ±1.6477	81.47	43.4667 ±1.8113	35.4667 ±1.7481	77.6000 ±1.5234

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	ngl	test	t/s
Dolphin3.0-Mistral-24B-Q4_K_M	12.95 GiB	23.57 B	CUDA	12	pp512	164.39 ± 0.20
Dolphin3.0-Mistral-24B-Q4_K_M	12.95 GiB	23.57 B	CUDA	12	tg128	4.71 ± 0.06
Dolphin3.0-Mistral-24B-Q4_K_M	12.95 GiB	23.57 B	CUDA	12	pp1024+tg1024	8.87 ± 0.04
Dolphin3.0-Mistral-24B-Q4_K_M-bartowski	13.34 GiB	23.57 B	CUDA	12	pp512	162.55 ± 0.47
Dolphin3.0-Mistral-24B-Q4_K_M-bartowski	13.34 GiB	23.57 B	CUDA	12	tg128	4.57 ± 0.03
Dolphin3.0-Mistral-24B-Q4_K_M-bartowski	13.34 GiB	23.57 B	CUDA	12	pp1024+tg1024	8.62 ± 0.03

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 orignal 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 gguf file format.