obliteratus

Removing refusals from Llama, Mistral, or MoE variants without retraining is achieved by identifying and projecting out refusal directions directly from the model weights. This method enables permissive behavior while bypassing fine-tuning.

Abliteration methods like surgical, nuclear, basic, and advanced differ in the strength and precision of refusal removal from LLM models. Choosing the right method tailors the safety and flexibility tradeoffs during mechanistic interpretability research.

Refusal removal works with Mixture of Experts (MoE) variants across both CPU and GPU setups. The process extracts refusal directions and validates the edited model using a compatible Python environment and the obliteratus CLI.

Analyzing refusal directions before removing them from a model is done using 28 analysis modules that evaluate mechanistic interpretability. These modules provide templates and telemetry to ensure reproducibility when testing guardrail behavior in open-weight LLMs.

Limitations of weight projection for refusal removal include potential risks to model coherence and the need to evaluate risk/benefit tradeoffs in model editing. Users must configure directions and quantization carefully to avoid degrading the target model.