What it is
A Python toolkit that implements abliteration: it probes a language model's hidden states, extracts the directions tied to refusals using methods such as SVD, PCA, and sparse autoencoders, and projects them out of the weights without retraining. It also runs ablation studies across layers and heads, checks perplexity and coherence afterward, and saves the edited model with metadata. Alignment researchers, red-teamers, and people running local models use it through a CLI, a Python API, a Colab notebook, or a Gradio interface on Hugging Face Spaces.
