eaddario/Dolphin-Mistral-24B-Venice-Edition-pruned-GGUF

Experimental layer-wise + pruned (layers 4 and 5) quantization of cognitivecomputations/Dolphin-Mistral-24B-Venice-Edition

Using LLaMA C++ release b5770 for quantization.

Original model: cognitivecomputations/Dolphin-Mistral-24B-Venice-Edition

From the original model creators:

Discord: https://discord.gg/h3K4XGj2RH
Website: https://dphn.ai
Twitter: https://x.com/dphnAI

What is Dolphin Mistral 24B Venice Edition?

Dolphin Mistral 24B Venice Edition is a collaborative project we undertook with Venice.ai with the goal of creating the most uncensored version of Mistral 24B for use within the Venice ecosystem.

Dolphin Mistral 24B Venice Edition is now live on https://venice.ai/ as “Venice Uncensored,” the new default model for all Venice users.

Dolphin aims to be a general purpose 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...

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 a custom version of llama-imatrix and llama-quantize to identify influential tensors, quantize the most important layers to higher bit precision and the less important to lower bits, and remove (prune) one or more layers. 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, and Xin Men's et al ShortGPT: Layers in Large Language Models are More Redundant Than You Expect

As of version b5125, llama-quantize can perform tensor-wide quantization (TWQ), whereby user-defined tensors are quantized at a specific level, or perform layer-wise quantization (LWQ) by selecting different quantization types per tensor/layer. For example, --tensor-type attn_v=q6_k will quantize all Attention Value tensors at q6_k (TWQ), and --tensor-type "\.([0-9]|1[01257]|31)\.attn_k=q4_k" will quantize Attention Key tensors on layers 0 to 9, 10, 11, 12, 15, 17 and 31 at q4_k, leaving the remaining layers at their default value (LWQ).

As of version b5740, llama-quantize can also prune models during quantisation by providing a comma-separated list in the --prune-layers command line option. The pruning operation will renumber remaining layers to avoid gaps in the sequence, update the relevant model metadata and, if an imatrix is available, it will use the correct importance score vector. This option can be used alongside --tensor-type to perform tensor/layer-wise quantization on selected tensor types, whilst at the same time pruning others. For example:

llama-quantize --tensor-type attn=q6_k --prune-layers 3,7,11 --imatrix imatrix.dat model-f32.gguf model-q4_k_m.gguf q4_k_m

An enhanced version of llama-imatrix generates useful statistics to guide the tensor and layer selection process. --show-statistics will display:

• Σ(Act²): the sum of all squared activations over the tensor (i.e. the Importance Scores)
• Min & Max: minimum and maximum squared activation values
• μ & σ: activations' mean and standard deviation
• % Active: proportion of elements whose average squared activation exceeds a very small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference
• N: number of squared activations in the tensor
• Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement)
• E (norm): Normalized entropy.
• ZD Score: z-score distribution as described in 3.1 Layer Importance Scores in the Layer-Wise Quantization paper
• CosSim: cosine similarity between same type tensors with respect to the previous layer (i.e. blk.7.attn_k and blk.6.attn_k)

Please note that statistics are calculated for each individual tensor and should be used to compare between tensors of the same type only. For example, assuming that attn_k in layer 10 has a higher influence during inference than attn_k in layer 7 because its Σ(Act²) is larger makes sense, whilst concluding the same between attn_k and ffn_down does not.

There’s a pull request to merge these changes back into the core llama.cpp project. This may or may not ever happen so, until then, the modified version will be available on GitHub.

For testing and comparison I use models produced by Unsloth (Daniel and Michael Han do some really advanced level stuff!) and Bartowski (see credits below) but if they don't provide versions of the required model, 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 modelled, 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 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 the enhanced version of llama-imatrix
5. Select an appropriate quant level for each tensor and quantize/prune the model using llama-quantize. In this model's case, layers 4 and 5 have been pruned
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	Bartowski	Repo	Shrinkage
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ3_M	10.7	9.6	10.3%
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ3_S	9.9	9.3	6.2%
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ4_NL	13.5	11.6	14.1%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_L	12.4	10.8	12.9%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_M	11.5	9.9	13.9%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_S	10.4	8.9	14.4%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M	14.3	12.4	13.3%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_S	13.5	11.7	13.3%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q5_K_M	16.8	14.3	14.9%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q5_K_S	16.3	13.9	14.7%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q6_K	19.7	16.8	14.7%
Dolphin-Mistral-24B-Venice-Edition-pruned-Q8_0	25.1	21.9	12.7%

Perplexity and KL Divergence scores

Model	μPPL	𝜌PPL	μKLD	RMS Δp
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ3_M	20.379006 ±0.160275	73.93%	1.290608 ±0.004304	37.928 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ3_S	21.165413 ±0.164512	73.80%	1.340446 ±0.004301	38.586 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ4_NL	18.783744 ±0.146959	74.79%	1.199318 ±0.004258	36.745 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_L	19.313300 ±0.150799	74.61%	1.248712 ±0.004216	37.260 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_M	18.723777 ±0.145380	75.90%	1.226150 ±0.004006	36.807 ±0.087
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_S	19.765437 ±0.153182	74.13%	1.295119 ±0.004177	38.004 ±0.087
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M	18.556910 ±0.145472	74.92%	1.187728 ±0.004237	36.521 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M-bartowski	6.304728 ±0.042418	99.60%	0.016941 ±0.000138	4.031 ±0.037
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_S	18.663517 ±0.146425	74.87%	1.192878 ±0.004250	36.598 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q5_K_M	18.174846 ±0.142320	75.14%	1.159685 ±0.004238	36.214 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q5_K_S	18.199918 ±0.142513	75.20%	1.160040 ±0.004229	36.220 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q6_K	18.213825 ±0.142965	75.05%	1.158026 ±0.004262	36.219 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-Q8_0	18.203515 ±0.142826	75.02%	1.158351 ±0.004265	36.227 ±0.088
Dolphin-Mistral-24B-Venice-Edition-pruned-F16	6.180577 ±0.041038	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 appropriate links: HellaSwag, ARC, MMLU, Truthful QA and WinoGrande

Model	ARC	HellaSwag	MMLU	Truthful QA	WinoGrande	Avg Score
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ3_M	65.6000 ±1.7358	79.60	42.9333 ±1.8086	38.4000 ±1.7771	72.4000 ±1.6334	59.79
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ3_S	64.9333 ±1.7436	79.87	42.0000 ±1.8034	38.0000 ±1.7736	72.5333 ±1.6309	59.47
Dolphin-Mistral-24B-Venice-Edition-pruned-IQ4_NL	68.4000 ±1.6988	80.66	44.9333 ±1.8176	38.1333 ±1.7748	74.4000 ±1.5947	61.31
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_L	67.2000 ±1.7155	80.27	43.2000 ±1.8100	39.6000 ±1.7870	72.9333 ±1.6235	60.64
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_M	66.6667 ±1.7225	80.67	43.8667 ±1.8132	39.4667 ±1.7860	72.2667 ±1.6358	60.59
Dolphin-Mistral-24B-Venice-Edition-pruned-Q3_K_S	66.2667 ±1.7276	78.93	43.7333 ±1.8126	38.1333 ±1.7748	72.8000 ±1.6260	59.97
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M	68.0000 ±1.7045	80.93	45.2000 ±1.8185	36.6667 ±1.7608	72.1333 ±1.6382	60.59
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M-bartowski	69.8667 ±1.6766	84.27	45.3333 ±1.8190	37.6000 ±1.7699	80.2667 ±1.4542	63.47
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_S	67.0667 ±1.7172	81.07	45.2000 ±1.8185	36.2667 ±1.7567	72.0000 ±1.6406	60.32
Dolphin-Mistral-24B-Venice-Edition-pruned-Q5_K_M	67.0667 ±1.7172	81.73	44.5333 ±1.8160	37.8667 ±1.7724	73.8667 ±1.6054	61.01
Dolphin-Mistral-24B-Venice-Edition-pruned-Q5_K_S	67.3333 ±1.7137	81.47	44.2667 ±1.8149	38.6667 ±1.7794	74.2667 ±1.5974	61.20
Dolphin-Mistral-24B-Venice-Edition-pruned-Q6_K	67.4667 ±1.7119	81.07	44.5333 ±1.8160	39.6000 ±1.7870	73.8667 ±1.6054	61.31
Dolphin-Mistral-24B-Venice-Edition-pruned-Q8_0	68.1333 ±1.7026	81.33	44.9333 ±1.8176	38.2667 ±1.7759	74.4000 ±1.5947	61.41
Dolphin-Mistral-24B-Venice-Edition-pruned-F16	70.8000 ±1.6614	84.53	45.3333 ±1.8190	38.1333 ±1.7748	80.2667 ±1.4542	63.81

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
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M	11.53 GiB	22.46 B	Metal,BLAS	12	pp512	266.57 ±14.60
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M	11.53 GiB	22.46 B	Metal,BLAS	12	tg128	27.60 ±0.54
Dolphin-Mistral-24B-Venice-Edition-pruned-Q4_K_M	11.53 GiB	22.46 B	Metal,BLAS	12	pp1024+tg1024	41.55 ±2.99
Dolphin-Mistral-24B-Venice-Edition-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	pp512	253.67 ±2.26
Dolphin-Mistral-24B-Venice-Edition-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	tg128	27.69 ±0.46
Dolphin-Mistral-24B-Venice-Edition-Q4_K_M-bartowski	13.34 GiB	23.57 B	Metal,BLAS	12	pp1024+tg1024	45.54 ±0.18

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

LLaMa C++ has a large and vibrant community of contributors (~1,200 last time I checked) that actively maintain and extend its functionality, adding new models and architectures almost as fast as they appear (considering the breakneck speed at which the AI/ML field is advancing, this alone is a remarkable feat!), and whilst I'm grateful to each and everyone of them, I want to recognise three people in particular: Thank You! Colin Kealty for the many contributions and for being one of the best sources of high quality quantized models available on Hugging Face, and a really big Thank You! to Georgi Gerganov for his amazing work with llama.cpp and the ggml/gguf libraries, and Iwan Kawrakow for being one of the key authors behind the many quantisation algorithms and the imatrix functionality.