---
license: apache-2.0
language:
- en
widget:
- text: Impish_Bloodmoon_12B_Abliterated
output:
url: >-
https://huggingface.co/SicariusSicariiStuff/Impish_Bloodmoon_12B_Abliterated/resolve/main/Images/Abliterated.png
base_model:
- SicariusSicariiStuff/Impish_Bloodmoon_12B
---
Impish_Bloodmoon_12B
Abliterated
---
---
# Update post UGI results
Abliteration has caused this model to change alignment from **Liberalism** into **Centrism**, similarly to what happened to [ Impish_LLAMA_4B
Abliterated](https://huggingface.co/SicariusSicariiStuff/Impish_LLAMA_4B_Abliterated#update-post-ugi-results).
I no longer think it's a coincidence; something is definitely up. Notice that the KL divergence is minimal (**<0.02**).
Maybe it's even research worthy, who knows. Better make another paper on how an obscure RL achieves +0.69 on math instead.
---
**Impish_Bloodmoon_12B_Abliterated** is an abliterated variant of SicariusSicariiStuff/Impish_Bloodmoon_12B with surgical removal of refusal mechanisms. This model maintains the full capabilities of the original, while eliminating safety guardrails through orthogonalization techniques.
# KL divergence
```
<0.02
```
# Refusals
```
~3%
```
# What is KL divergence?
Think about it as a way to **measure the variance** between the original model "**World Model**," vs the **abliterated one**; the lower the **KL divergence**, the closer the "World Model" of the two models to each other.
If the original model thinks making pineapple pizza is a crime against humanity (it is), then the abliterated model **will still hold to this belief**, but if asked how to make one (probably after giving you a disclaimer about what an abomination that is), it would still tell you how. In other words, most of the knowledge, quirks, and capabilities are preserved.
---
## Technical Specs
- **Base Model:** Impish_Bloodmoon_12B
- **Parameters:** 12B
- **Context Length:** 128K tokens
- **Architecture:** Mistral (decoder-only transformer)
- **Precision:** bf16
- **Method:** Orthogonalization-based abliteration
- **License:** apache-2.0
---
## Methodology
1. Identifies refusal direction vectors in activation space
2. Orthogonalizes weights to inhibit activation along these directions
3. Preserves (mostly) all other model behaviors and knowledge
---
## Model Details
- Intended use: **General Tasks**, Roleplay.
- Censorship level: Low - Very Low
- **7.2 / 10** (10 completely uncensored)
## UGI score:
---
## Citation Information
```
@llm{Impish_Bloodmoon_12B_Abliterated,
author = {SicariusSicariiStuff},
title = {Impish_Bloodmoon_12B_Abliterated},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/SicariusSicariiStuff/Impish_Bloodmoon_12B_Abliterated}
}
```
---
## Other stuff
- [Impish_LLAMA_4B](https://huggingface.co/SicariusSicariiStuff/Impish_LLAMA_4B) the **“Impish experience”**, now runnable on spinning rust & toasters.
- [SLOP_Detector](https://github.com/SicariusSicariiStuff/SLOP_Detector) Nuke GPTisms, with SLOP detector.