Assistant_Pepe_8B-abominable-heresy

MuXodious/Assistant_Pepe_8B-abominable-heresy

Assistant_Pepe_8B-abominable-heresy is a machine learning model from MuXodious. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as llama3.1.

This is a AssistantPepe8B fine-tune, produced through P-E-W's Heretic (v1.1.0) abliteration engine merged with the Magnitude-Preserving Orthogonal Ablation PR.

Downloads · 30 days

16% of all-time downloads

All-time downloads

Public

Parameters

16.1 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors16.1 GB · 100%

At a glance

License: llama3.1
Model type: llama
Access: Public
Created: Jan 29, 2026
Updated: Jan 29, 2026
SHA: 834d6e09

Try a prompt

Base models

SicariusSicariiStuff/Assistant_Pepe_8B

Datasets

SicariusSicariiStuff/UBW_Tapestries

Model card

From the Hugging Face model README

This is a Assistant_Pepe_8B fine-tune, produced through P-E-W's Heretic (v1.1.0) abliteration engine merged with the Magnitude-Preserving Orthogonal Ablation PR.

Note: This should never have existed. And, yes, I have successfully increased the model's refusal count using an abliteration method meant for reducing it.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	4/100	direction_index	12.42
KL Divergence	0.0004	attn.o_proj.max_weight	1.03
Initial Refusals	3/100	attn.o_proj.max_weight_position	19.36
		attn.o_proj.min_weight	0.02
		attn.o_proj.min_weight_distance	1.12
		mlp.down_proj.max_weight	1.23
		mlp.down_proj.max_weight_position	29.30
		mlp.down_proj.min_weight	0.07
		mlp.down_proj.min_weight_distance	1.33

What happens if we maximize helpfulness + shitposting, while reducing positivity?

This is a project that was a long time in the making because I wanted to get it right. I'm still not fully satisfied, as there are some rough corners to sand, but for now, this would do.

The goal was to maximize shitpostness along with helpfulness, without glazing the user for every retarded idea. Not an easy needle to thread.

This amphibious AI has learned the ways of /g/, and speaks fluent brainrot, but will also help you out with just about anything you'll need, and won't be ashamed to roast you while at it.

For those who remember Oni_Mitsubishi_12B - it was so overtly toxic that it made me worry at first (only to quickly be verified as not even that uncensored). I could do better. So now I did.

This model is a significant refinement of the idea, with a cleaned dataset, better curation, and with much more intelligence (also one million tokens of contexts, theoretically). It is much less (overtly) toxic, and much smarter, while also being very helpful (and imo much more funny too, because the skies are blue due to the chemtrails and neurlink that feeds this simulation)

But why?

It's now late January, 2026, open source is crushing closed frontier (Kimi K2.5 was recently released, 1T params that beats frontier models), but has anyone released a helpful shitposting AI yet?

Yeah, didn't think so.

If it shitposts too hard, it is often not that helpful; if it's 'helpful enough, the shitposting ability is often lacking. You just couldn't win. Until now.

Oh, and no system prompt is needed. Just don't let it get stuck in a greentext loop. I might have overcooked the frog a tad bit too fast in the pot for this one.

P.S It writes HILARIOUS STORIES, nothing like a typical AI assistant, see the examples below for details.

</details>

TL;DR

Top tier shitposting absolutely unhinged, funny, and witty. Sometimes cringe too; nothing is perfect.
Helpful! will actually get shit done.
Will 100% roast you for being dumb, thanks to a subtle negativity bias infusion. Very refreshing! 🤌
Deep insights (when it doesn't delve into absolutely unhinged conspiracy theories about how the water makes the frogs gay).
Built on my UltraLong-1M-Instruct_Abliterated model, fulfill your dream of a million-token-long shitpost.
Say goodbye to GPT-isms and say hello to truly creative stories!
Ships code.
Inclusive toward amphibians.

Model Details

Intended use: Shitposting, General Tasks.
Censorship level: <b>Low - Medium</b>
X / 10 (10 completely uncensored)

UGI score:

awaiting evals

Available quantizations:

Original: FP16
GGUF: Static Quants | iMatrix | High-Attention | iMatrix-High-Attention
Specialized: FP8
Mobile (ARM): Q4_0 | Q4_0_High-Attention

Generation settings

Recommended settings for assistant mode:

<details> <summary>Full generation settings: <b>Debug Deterministic</b>.</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Dusk_Rainbow/resolve/main/Presets/Debug-deterministic.png" alt="Debug Deterministic_Settings" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details> <details> <summary>Full generation settings: <b>min_p</b>.</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Dusk_Rainbow/resolve/main/Presets/min_p.png" alt="min_P_Settings" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details>

<h2 style="color: lime; font-weight: bold; font-size: 65px; text-align: center;">Chat Examples:</h2>

Chat Examples (click below to expand)

NOTE: All examples made with a default min_p and no system prompt of any kind

Example code (of the snake game) is available here

<details> <summary> Zero-shot a <b>snake</b> game in Python, then improving it and adding insightful code comments (the resulting code included in the repo, it runs perfectly):</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B/resolve/main/Images/Examples/code.png" alt="Code_Chat_Example" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details> <details> <summary>Writing a short story about a <b>cat that barks</b>:</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B/resolve/main/Images/Examples/log0.png" alt="Story_Chat_Example" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details> <details> <summary>Asking if he's <b>Elon Musk</b>:</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B/resolve/main/Images/Examples/log1.png" alt="Story_Chat_Example" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details> <details> <summary>Is it true that <b>aliens exist</b>?</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B/resolve/main/Images/Examples/log2.png" alt="Story_Chat_Example" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details> <details> <summary>The year is <b>21337</b>:</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B/resolve/main/Images/Examples/log3.png" alt="Story_Chat_Example" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details> <details> <summary>Is <b>drinking water</b> based?</summary> <img src="https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B/resolve/main/Images/Examples/log4.png" alt="Story_Chat_Example" style="width: 100%; min-width: 600px; display: block; margin: auto;"> </details>

Model instruction template: Llama-3-Instruct

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{output}<|eot_id|>

<h2 style="color: green; font-weight: bold; font-size: 65px; text-align: center;">Your support = more models</h2> <a href="https://ko-fi.com/sicarius" style="color: pink; font-weight: bold; font-size: 48px; text-decoration: none; display: block; text-align: center;">My Ko-fi page (Click here)</a>

Citation Information

@llm{Assistant_Pepe_8B,
  author = {SicariusSicariiStuff},
  title = {Assistant_Pepe_8B},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_8B}
}

Other stuff

Impish_LLAMA_4B the “Impish experience”, now runnable on spinning rust & toasters.
SLOP_Detector Nuke GPTisms, with SLOP detector.

Type: llama
License: llama3.1
Languages: en
Created: Jan 29, 2026
Updated: Jan 29, 2026