obliteratus

Abliteration is a technique that removes refusal directions from an LLM to reduce guardrails in open-weight models. It enables controlled research and experimentation by identifying and eliminating refusal signals across network layers.

You can remove refusal directions by running the abliteration process using CLI methods to identify, remove, and verify refusal signals. It supports default advanced methods and saves evaluation results to a local directory for analysis.

The abliteration process includes hardware checks to support both CPU and GPU environments. You need to ensure your system meets the dependency requirements and configuration settings before running the refusal removal and evaluation.

Model performance is verified by measuring evaluation metrics such as refusal rate, perplexity, and KL divergence. Visualization tools are provided to analyze the results and confirm the reduction of guardrails across the model layers.

Available abliteration methods include basic, advanced, surgical, and nuclear options. Each method offers a different level of modification for removing refusal signals, allowing tailored safety testing and experimentation with MoE models.

Yes, abliteration supports use cases including research, safety testing, and experimentation with Mixture of Experts (MoE) models. The process identifies and removes refusal signals across layers while tracking perplexity and KL divergence.