obliteratus

To remove refusals from open-weight LLMs without retraining, you can surgically erase refusal directions in model weights using mechanistic interpretability techniques. This process, known as abliteration, modifies weights directly while preserving reasoning capabilities.

Abliteration is the process of surgically erasing refusal directions from LLM weights. It uses mechanistic interpretability techniques like SVD, whitened SVD, and diff-in-means to identify and remove refusal behaviors while aiming to preserve the model's reasoning capabilities.

Yes, abliteration requires CLI tooling and dependencies such as torch, transformers, and safetensors. The process outputs a standard HuggingFace-compatible model directory, allowing you to load and run the uncensored model using standard transformers pipelines.

Yes, abliteration can be applied to MoE models. The skill includes configurable diagnostics and a rich analysis suite specifically designed for mechanistic interpretability across 116 model presets, ensuring refusal removal works across different architectures.

While weight projection removes refusals without retraining, the primary limitation is the risk of degrading reasoning capabilities. The skill includes 28 analysis modules and telemetry-driven recommendations to guide safe experimentation and monitor alignment imprints across compute tiers.

Yes, obliteration supports multiple methods for erasing refusal directions, including LEACE and SAE decomposition. It offers 9 CLI methods ranging from basic to nuclear approaches, allowing researchers to study refusal mechanics and evaluate safety controls.