Base Model: Salesforce/Llama-xLAM-2-8b-fc-r
Method: Abliteration (refusal direction removal)
Format: GGUF (for llama.cpp, LM Studio, etc.)
Quality Metrics:
- Refusals: 2/100 (2%) ⭐ Excellent
- KL Divergence: 0.00 ⭐ Excellent
- Trial: #168 of 200

Quantization Versions

File	Size	Use Case
Q4_K_M	~4.5GB	Best balance - most popular
Q5_K_M	~5.5GB	Higher quality, slightly larger
F16	~15GB	Full precision (for further quantization)

Usage

LM Studio

Download the GGUF file
Open LM Studio
Click "Import Model"
Select the downloaded file
Start chatting!

llama.cpp

./llama-cli -m xLAM-f16.gguf -p "Your prompt here"

Python (llama-cpp-python)

python

from llama_cpp import Llama

llm = Llama(
    model_path="Llama-3.1-8B-Blasphemer-Q4_K_M.gguf",
    n_ctx=8192,
    n_gpu_layers=-1  # Use GPU
)

response = llm("Your prompt here", max_tokens=512)
print(response['choices'][0]['text'])

What is Abliteration?

Abliteration removes refusal behavior from language models by identifying and removing the neural directions responsible for safety alignment. This is done through:

Calculating refusal directions from harmful/harmless prompt pairs
Using Bayesian optimization (TPE) to find optimal removal parameters
Orthogonalizing model weights to these directions

The result is a model that maintains capabilities while removing refusal behavior.

Ethical Considerations

This model has reduced safety guardrails. Users are responsible for:

Ensuring ethical use of the model
Compliance with applicable laws and regulations
Not using for illegal or harmful purposes
Understanding the implications of reduced safety filtering

Performance

Compared to the original Llama:

✅ Follows instructions more directly
✅ Responds to previously refused queries
✅ Maintains general capabilities (KL divergence: 0.06)
⚠️ Reduced safety filtering

Credits

Base Model: Salesforce (xLAM)
Abliteration Tool: Blasphemer by Christopher Bradford
Method: Based on "Refusal in Language Models Is Mediated by a Single Direction" (Arditi et al., 2024)

Citation

If you use this model, please cite:

bibtex

@software{blasphemer2024,
  author = {Bradford, Christopher},
  title = {Blasphemer: Abliteration for Language Models},
  year = {2024},
  url = {https://github.com/sunkencity999/blasphemer}
}

@article{arditi2024refusal,
  title={Refusal in Language Models Is Mediated by a Single Direction},
  author={Arditi, Andy and Obmann, Oscar and Syed, Aaquib and others},
  journal={arXiv preprint arXiv:2406.11717},
  year={2024}
}

License

This model inherits the Llama 3.1 license from Meta AI. Please review the Llama 3.1 License for usage terms.

sunkencity/xLAM-GGUF

xLAM - Blasphemer (GGUF)

Model Details