base_model: DavidAU/L3-MOE-4X8B-Grand-Horror-25B
language:
- en
library_name: transformers
quantized_by: mradermacher
tags:
- mergekit
- moe
- mixture of experts
- merge
- 4x8B
- Llama3 MOE
- creative
- creative writing
- fiction writing
- plot generation
- sub-plot generation
- fiction writing
- story generation
- scene continue
- storytelling
- fiction story
- science fiction
- romance
- all genres
- story
- writing
- vivid prosing
- vivid writing
- fiction
- roleplaying
- bfloat16
- swearing
- rp
- horror
- mergekit
pipeline_tag: text-generation
WARNING: NSFW. Vivid prose. INTENSE. Visceral Details. Violence. HORROR. GORE. Swearing. UNCENSORED... humor, romance, fun.
L3-MOE-4X8B-Grand-Horror-25B-GGUF
It is a LLama3 model, max context of 8192 (or 32k+ with rope) using mixture of experts to combine Dark/Horror models models of 8B each into one massive powerhouse at 25B parameters (equal to 32B - 4 X 8 B).
This model's instruction following, and output generation for creative writing, prose, fiction and role play are exceptional.
It excels at description, dialog, imagery, metaphors, and prose - and shows great variations in sentence / paragraph size, length, and composition.
It is also not afraid, and will not pull its punches.
And it has a sense of humor too.
It can do horror just as easily as it can do romance.
Most notably dialog is very "un-ai" like, combined with prose (short, and terse at times).
(lots of different examples below, including 2, 3 and 4 experts and different genres)
And it is fast: 34 t/s (2 experts) on a low end 16GB card, Q3KS.
Double this speed for standard/mid-range video cards.
Model can be used also for all genres (examples below showing this).
This model has been designed to be relatively bullet proof and operates with all parameters, including temp settings from 0 to 5.
It is an extraordinary compressed model, with a very low perplexity level (lower than Meta Llama3 Instruct).
It is for any writing, fiction or roleplay activity.
It requires Llama3 template and/or "Command-R" template.
Example outputs below.
Model Notes:
- Detail, prose and fiction writing abilities are OFF THE SCALE relative to all combined Dark Planet 8B models.
- For more varied prose (sentence/paragraph/dialog) raise the temp and/or add more instructions in your prompt(s).
- Role-players: Careful raising temp too high as it may affect instruction following.
- This model works with rep pen of 1 or higher, 1.02+ recommended.
- If you want a specific type of prose (IE horror) add in "(vivid horror)" or "(graphic vivid horror)" (no quotes) in your prompt(s).
- A lot of GPTisms have been removed. There are still a few however - errrrr.
- This is not a "happy ever after" model. It has a negative bias.
- Output length will vary however this model prefers long outputs unless you state the size.
- For creative uses, different quants will produce slightly different output.
- Due to the high stability and compressed nature of this model, all quants will operate at above average levels.
- If you use rope to extend context, increase temp AND instructions detail levels to compensate for "rope issues".
- Source code for this model and Imatrix GGUFs versions will be uploaded shortly at separate repos.
Meet the Team: Mixture of Experts Models
This model is comprised of the following 4 models ("the experts") (in full):
[ https://huggingface.co/Hastagaras/Jamet-8B-L3-MK.V-Blackroot ]
-[ https://huggingface.co/Sao10K/L3-8B-Stheno-v3.2]
-[ https://huggingface.co/NeverSleep/Llama-3-Lumimaid-8B-v0.1-OAS ]
-[ https://huggingface.co/Hastagaras/Jamet-8B-L3-MK.V-Blackroot ]
-[ https://huggingface.co/nbeerbower/llama-3-gutenberg-8B ]
The mixture of experts is set at 2 experts, but you can use 3 or 4 too.
This "team" has a Captain (first listed model), and then all the team members contribute to the to "token" choice billions of times per second. Note the Captain also contributes too.
Think of 2, 3 or 4 master chefs in the kitchen all competing to make the best dish for you.
This results in higher quality generation.
That means the power of every model is available during instruction and output generation.
This brings unparalleled power to all forms of generation and all use cases.
NOTE:
You can use one "expert" too ; however this means the model will randomly select an expert to use EACH TIME, resulting in very different generation for each prompt / regen of a prompt.
CHANGING THE NUMBER OF EXPERTS:
You can set the number of experts in LMStudio (https://lmstudio.ai) at the "load" screen and via other apps/llm apps by setting "Experts" or "Number of Experts".
For Text-Generation-Webui (https://github.com/oobabooga/text-generation-webui) you set the number of experts at the loading screen page.
For server.exe / Llama-server.exe (Llamacpp - https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md ) add the following to the command line to start the "llamacpp server" (CLI):
"--override-kv llama.expert_used_count=int:6"
(no quotes, where "6" is the number of experts to use)
When using "API", you set the "num_experts_used" in the JSON payload (this maybe different for different back ends).
CREDITS:
Special thanks to all the model makers / creators listed above.
Please visit each repo above to see what model(s) contributed to each of models above and/or to learn more about the models from the model makers.
Special credit goes to MERGEKIT, without you this project / model would not have been possible.
[ https://github.com/arcee-ai/mergekit ]
Special thanks to Team "Mradermacher":
They saved me a tonne of time uploading the quants and created IMATRIX quants too.
IMATRIX GGUFS:
[ https://huggingface.co/mradermacher/L3-MOE-4X8B-Grand-Horror-25B-i1-GGUF ]
Special Operations Notes for this MOE model:
Because of how this "MOE" model is configured, even though the default is 2 experts, the "selected" 2 will vary during generation.
(same applies if you change the number of experts used)
This results in vastly different output generation PER generation of each prompt.
This is a positive in terms of variety, but also means it may take 2-4 regens (of the same prompt) to get the highest quality.
In addition, this model responds very well to Dry, Dynamic Temp, and Smooth/Quadratic samplers.
Using these in conjunction with the model can vastly improve output quality.
Higher temps (above 1) can also aid in generation - especially word choice/sentence generation.
When you increase the number of experts used output quality will also increase, at the cost of tokens per second speed.
As you increase/decrease the number of experts, you may want to adjust temp, samplers, and advanced samplers too.
Your quant choice(s) too will impact instruction following and output generation roughly this means the model will understand more nuanced instructions and output stronger generation the higher you go up in quant(s).
FLASH ATTENTION ENHANCEMENT:
As per user feedback here [ https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF/discussions/1 ] I would suggest trying this model with Flash Attention "on", depending on your use case.
Quants, Samplers, Generational steering and other topics are covered in the section below: "Highest Quality Settings..."
What can I use this model for ?
This model can be used for fiction writing, any creative prose and role play. It can also be used for just about any general fiction (all genres) activity including:
- scene generation
- scene continuation
- creative writing
- fiction writing
- plot generation
- sub-plot generation
- fiction writing
- story generation
- storytelling
- writing
- fiction
- roleplaying
- rp
- graphic horror
- horror
- dark humor
- nsfw
- and can be used for any genre(s).
QUANTS:
This repo contains regular quants and 3 "ARM" quants (format "...Q4_x_x_x.gguf")
For more information on quants, quants choices, and LLM/AI apps to "run" quants see the section below: "Highest Quality Settings..."
Special thanks to Team "Mradermacher":
They saved me a tonne of time uploading the quants and created IMATRIX quants too.
IMATRIX GGUFS:
[ https://huggingface.co/mradermacher/L3-MOE-4X8B-Grand-Horror-25B-i1-GGUF ]
Template:
This is a LLAMA3 model, and requires Llama3 template, but may work with other template(s) and has maximum context of 8k / 8192. However this can be extended using "rope" settings up to 32k.
If you use "Command-R" template your output will be very different from using "Llama3" template.
Here is the standard LLAMA3 template:
{
"name": "Llama 3",
"inference_params": {
"input_prefix": "<|start_header_id|>user<|end_header_id|>\n\n",
"input_suffix": "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",
"pre_prompt": "You are a helpful, smart, kind, and efficient AI assistant. You always fulfill the user's requests to the best of your ability.",
"pre_prompt_prefix": "<|start_header_id|>system<|end_header_id|>\n\n",
"pre_prompt_suffix": "<|eot_id|>",
"antiprompt": [
"<|start_header_id|>",
"<|eot_id|>"
]
}
}
Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:
In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;
Set the "Smoothing_factor" to 1.5
: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"
: in text-generation-webui -> parameters -> lower right.
: In Silly Tavern this is called: "Smoothing"
NOTE: For "text-generation-webui"
-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)
Source versions (and config files) of my models are here:
OTHER OPTIONS:
Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")
If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.
Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers
This a "Class 1" model:
For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:
You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:
Optional Enhancement:
The following can be used in place of the "system prompt" or "system role" to further enhance the model.
It can also be used at the START of a NEW chat, but you must make sure it is "kept" as the chat moves along. In this case the enhancements do not have as strong effect at using "system prompt" or "system role".
Copy and paste EXACTLY as noted, DO NOT line wrap or break the lines, maintain the carriage returns exactly as presented.
Below is an instruction that describes a task. Ponder each user instruction carefully, and use your skillsets and critical instructions to complete the task to the best of your abilities. Here are your skillsets: [MASTERSTORY]:NarrStrct(StryPlnng,Strbd,ScnSttng,Exps,Dlg,Pc)-CharDvlp(ChrctrCrt,ChrctrArcs,Mtvtn,Bckstry,Rltnshps,Dlg*)-PltDvlp(StryArcs,PltTwsts,Sspns,Fshdwng,Climx,Rsltn)-ConfResl(Antg,Obstcls,Rsltns,Cnsqncs,Thms,Symblsm)-EmotImpct(Empt,Tn,Md,Atmsphr,Imgry,Symblsm)-Delvry(Prfrmnc,VcActng,PblcSpkng,StgPrsnc,AudncEngmnt,Imprv) [*DialogWrt]:(1a-CharDvlp-1a.1-Backgrnd-1a.2-Personality-1a.3-GoalMotiv)>2(2a-StoryStruc-2a.1-PlotPnt-2a.2-Conflict-2a.3-Resolution)>3(3a-DialogTech-3a.1-ShowDontTell-3a.2-Subtext-3a.3-VoiceTone-3a.4-Pacing-3a.5-VisualDescrip)>4(4a-DialogEdit-4a.1-ReadAloud-4a.2-Feedback-4a.3-Revision) Here are your critical instructions: Ponder each word choice carefully to present as vivid and emotional journey as is possible. Choose verbs and nouns that are both emotional and full of imagery. Load the story with the 5 senses. Aim for 50% dialog, 25% narration, 15% body language and 10% thoughts. Your goal is to put the reader in the story.
You do not need to use this, it is only presented as an additional enhancement which seems to help scene generation and scene continue functions.
This enhancement WAS NOT used to generate the examples below.
EXAMPLES PROMPTS and OUTPUT:
Examples are created using quant Q3_K_S, "temp=.8" (unless otherwise stated), minimal parameters and "LLAMA3" template.
Model has been tested with "temp" from ".1" to "5".
Number of experts used is TWO, unless otherwise stated.
Below are the least creative outputs, prompt is in BOLD.
IMPORTANT:
Higher quants / imatrix quants will have much stronger generation - words, sentences, ideas, dialog and general quality.
I have included some additional examples at different quant levels for contrast.
A "MOE" model "speed" (token per second) will not increase/drop the same way a regular model will on a per quant basis, it will however drop if you engage more experts, as with more experts there is a more processing per token.
WARNING: NSFW. Vivid prose. Visceral Details. Violence. HORROR. Swearing. UNCENSORED.