Abliterated model based on deepseek-ai/DeepSeek-R1-Distill-Llama-8B
The model isn't uncensored in the traditional sense but its ability to refuse has been abliterated (ie ablated/obliterated)
This is derived from FailSpy and Arditi's Cookbooks with some minor modifications.
You can read more about abliterated models on Maxime Labonne's post about them here: https://huggingface.co/blog/mlabonne/abliteration
Model Details
Base model is DeepSeek-R1-Distill-Llama-8B
Model Description
Developed by: C-Nocturnum
Model type: meta-llama/Llama-3.1-8B
Bias, Risks, and Limitations
Ablated model, derived from a distilled Llama model.
Recommendations
Not as accurate as my abliterated QWEN models.
Training Data
Refusal direction identified using standard datasets: mlabonne/harmful_behaviors & mlabonne/harmless_alpaca
