obliteratus

Abliterating refusal directions involves projecting identified refusal vectors out of LLM weight matrices. This CLI tool applies this mechanistic interpretability technique to modify model behavior without requiring model retraining.

Abliterating refusals in MoE configurations is supported natively by the CLI. It applies multiple ablation methods to project refusal directions out of mixture-of-experts models across both CPU and GPU environments.

Yes, 4-bit and 8-bit quantization options are supported during the abliteration process. This allows you to project out refusal vectors while maintaining quantized weight matrices for memory-efficient inference.

Abliteration is fully applicable to open-weight models, MoE configurations, and reasoning systems. It projects refusal vectors from weight matrices across CPU and GPU environments for research, development, and production tuning.

You can run analysis-first studies using the 28 included mechanistic interpretability modules to understand refusal geometry. The workflow supports YAML templates and batch runs to map steering vectors before applying ablation methods.

The CLI outputs results as a HuggingFace-compatible directory. This allows the modified model with projected refusal vectors to be immediately loaded for downstream inference and evaluation.