refusal-direction

The world's top open-weight AI model workshop. Measure, uncensor and audit: refusal removal for transformer language models, with the instruments to measure what the edit cost.

Automated, capability-preserving abliteration for open-weight LLMs — agent-native (MCP server). Clean-room MIT implementation.

Model-agnostic abliteration toolkit for HuggingFace causal LMs: compute and remove refusal directions via weight orthogonalization or runtime hooks, with KL-aware Optuna optimization, quantized-model bundles, and a local web UI.

Mechanistic study of the refusal direction across base, instruction-tuned, and reasoning-distilled Qwen2.5-1.5B variants: extraction, ablation, transplant, and phase-aware analysis.

Gemma 4 abliteration and mechanistic interpretability lab for refusal-direction extraction, activation steering, weight editing, and rigorous LLM evaluation.

Remove refusal behaviour from an instruct LLM by ablating the refusal direction (Arditi et al., NeurIPS 2024). Self-contained: prompt sets included; Qwen2.5-1.5B-Instruct by default.

Reference-free detection of abliteration in open-source LLMs — tell whether a model was censored, fine-tuned, or silently uncensored, without the original model.

Mechanistic interpretability toolkit for LLM refusal: locate and ablate the linear refusal direction (Arditi et al. 2024) without retraining. Diff-Means, whitened SVD, COSMIC layer selection, reversible steering vectors. PyTorch + Transformers.