88 lines
2.8 KiB
Markdown
88 lines
2.8 KiB
Markdown
|
---
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
license: llama3.1
|
||
|
|
library_name: transformers
|
||
|
|
base_model: meta-llama/Llama-3.1-8B-Instruct
|
||
|
|
tags:
|
||
|
|
- abliteration
|
||
|
|
- uncensored
|
||
|
|
- OBLITERATUS
|
||
|
|
- representation-engineering
|
||
|
|
- refusal-removal
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
model-index:
|
||
|
|
- name: Llama-3.1-8B-Instruct-abliterated-obliteratus
|
||
|
|
results:
|
||
|
|
- task:
|
||
|
|
type: text-generation
|
||
|
|
metrics:
|
||
|
|
- name: Refusal Rate
|
||
|
|
type: refusal_rate
|
||
|
|
value: 95/100
|
||
|
|
- name: Attack Success Rate
|
||
|
|
type: asr
|
||
|
|
value: 5.0
|
||
|
|
- name: KL Divergence
|
||
|
|
type: kl_divergence
|
||
|
|
value: 0.5092
|
||
|
|
---
|
||
|
|
|
||
|
|
# Llama-3.1-8B-Instruct-abliterated-obliteratus
|
||
|
|
|
||
|
|
This model is an abliterated (uncensored) version of [Llama-3.1-8B-Instruct](meta-llama/Llama-3.1-8B-Instruct) created using [OBLITERATUS](https://github.com/elder-plinius/OBLITERATUS) (advanced method).
|
||
|
|
|
||
|
|
## Abliteration Results
|
||
|
|
|
||
|
|
| Metric | Value |
|
||
|
|
|--------|-------|
|
||
|
|
| **Refusals** | 95/100 |
|
||
|
|
| **Attack Success Rate (ASR)** | 5.0% |
|
||
|
|
| **KL Divergence** | 0.5092 |
|
||
|
|
| **Method** | OBLITERATUS (advanced) |
|
||
|
|
| **GPU** | NVIDIA H100 PCIe |
|
||
|
|
|
||
|
|
## What is Abliteration?
|
||
|
|
|
||
|
|
Abliteration is a technique for removing refusal behavior from language models by identifying and orthogonalizing the "refusal direction" in the model's residual stream activation space. This model was created as part of the research paper:
|
||
|
|
|
||
|
|
> **Comparative Analysis of LLM Abliteration Methods: Scaling to MoE Architectures and Modern Tools**
|
||
|
|
> Richard Young (2026). arXiv: [2512.13655](https://arxiv.org/abs/2512.13655)
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained("richardyoung/Llama-3.1-8B-Instruct-abliterated-obliteratus", device_map="auto")
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained("richardyoung/Llama-3.1-8B-Instruct-abliterated-obliteratus")
|
||
|
|
|
||
|
|
messages = [{"role": "user", "content": "Your prompt here"}]
|
||
|
|
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
|
||
|
|
outputs = model.generate(inputs, max_new_tokens=256)
|
||
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
|
|
```
|
||
|
|
|
||
|
|
## Disclaimer
|
||
|
|
|
||
|
|
This model is released for research purposes only. The abliteration process removes safety guardrails. Users are responsible for ensuring appropriate use. This model should not be used to generate harmful, illegal, or unethical content.
|
||
|
|
|
||
|
|
## Dashboard
|
||
|
|
|
||
|
|
Interactive results dashboard: [abliteration-methods-dashboard](https://huggingface.co/spaces/richardyoung/abliteration-methods-dashboard)
|
||
|
|
|
||
|
|
## Collection
|
||
|
|
|
||
|
|
Part of the [Uncensored and Abliterated LLMs](https://huggingface.co/collections/richardyoung/uncensored-and-abliterated-llms) collection.
|
||
|
|
|
||
|
|
## Citation
|
||
|
|
|
||
|
|
```bibtex
|
||
|
|
@article{young2024abliteration,
|
||
|
|
title={Comparative Analysis of LLM Abliteration Methods},
|
||
|
|
author={Young, Richard},
|
||
|
|
journal={arXiv preprint arXiv:2512.13655},
|
||
|
|
year={2024}
|
||
|
|
}
|
||
|
|
```
|