Refusal and guardrail behaviors in modern LLMs are often implemented by geometric features in activation space rather than monolithic code paths. That means they can be located, measured, and—if desired—surgically projected out without full retraining. OBLITERATUS wraps that observation into a reproducible pipeline that makes the analysis-to-edit loop auditable and repeatable for research or deployment experimentation.
What Sets It Apart
- Analysis-informed pipeline: runs cross-layer and cone-geometry analyses before choosing how many and which directions to remove, so interventions are tailored to entanglement and self-repair risk (the "Ouroboros" effect). This reduces blind overprojection and helps preserve reasoning capabilities.
- Multiple extraction & intervention modes: implements diff-in-means, SVD, whitened SVD, sparse-autoencoder decompositions, norm-preserving projections, and reversible steering vectors — offering both permanent weight projection and inference-time steering trade-offs.
- Low-friction entry points: a zero-setup HuggingFace Space and a Colab notebook for experimentation, plus a CLI and Python API for reproducible studies and automated pipelines. Runs optionally contribute anonymized telemetry for community aggregation.
- Built-in verification & metrics: post-edit checks (perplexity, coherence, KL divergence, refusal-rate) and iterative re-probing to detect rotated residual guardrails.
Who It's For and Trade-offs
Great fit if you are a researcher or practitioner studying mechanistic interpretability, alignment geometry, or need to experiment with behavior-editing without retraining. It is useful for comparative experiments, ablation studies, and rapid prototyping of steering vs. permanent edits. Look elsewhere if you require enterprise-friendly licensing without AGPL constraints (project is AGPL-3.0) or if you need a formal safety review before deploying on public-facing services: removing guardrails carries ethical and legal risks and may violate third-party terms or safety requirements. Expect large-model runs to be I/O and verification-time dominated; quantization and correct GPU sizing are often necessary for bigger checkpoints.
Where It Sits Compared to Similar Tools
Compared to hook-and-probe libraries, the bundle emphasizes an end-to-end, analysis→distill→excise workflow with community telemetry and a UI for non-coders. It complements mechanistic toolkits (TransformerLens-style tracing) by offering ready-made extraction, projection, and steering implementations oriented at refusal directions rather than generic concept attribution.
