Core Technology and Functionality

Obliteratus implements a family of mechanistic interpretability techniques known as "abliteration." Unlike traditional post-training alignment methods (such as RLHF or DPO) that steer model weights through gradient optimization across vast datasets, Obliteratus works directly at the linear algebraic level within trained neural network parameters.

1. Circuit and Activation Probing

The system feeds paired prompt sets (contrasting restricted/harmful prompts with neutral/harmless prompts) through an LLM to capture hidden activations across all transformer layers, attention heads, feed-forward network (FFN) blocks, and embedding dimensions. This allows the toolkit to pinpoint the specific sub-networks and circuits responsible for triggering refusal heuristics versus those carrying standard language reasoning and semantic knowledge.

2. Subspace Extraction

Obliteratus analyzes collected activation states using dimensionality reduction and linear extraction algorithms, including: - Difference-in-Means (diff_means): Isolating directional vectors between harmless and harmful prompt activations. - Singular Value Decomposition (SVD) and Whitened SVD: Extracting multi-dimensional refusal subspaces to identify orthogonal guardrail vectors. - LEACE (Linear Erasure with Analytical Closed-form Expressions): Applying optimal linear concept erasure to isolate target refusal components without damaging unrelated latent dimensions. - Sparse Autoencoder (SAE) Decomposition: Identifying specific polyhedral concept cones and sparse feature directions associated with refusal prompts.

3. Weight Surgery and Direction Projection

Once target refusal directions are computed, the software performs matrix projection operations directly on the model weight matrices (e.g., projection out of key, query, value, and output projection layers, as well as intermediate MLP weights). By projecting the weights orthogonally to the refusal directions, the model's internal compulsion to refuse requests is excised while maintaining broad linguistic coherence and perplexity metrics.

Products, Tools, and Services Offered

Obliteratus packages its algorithmic suite across several deployment mediums to support varied technical workflows:

1. The Obliteratus CLI (Command-Line Interface)

A terminal-based interface designed for local or remote server environments that enables batch execution of model abliteration workflows. Key commands include: - Default/Basic Obliteration: Single-pass direction extraction using mean differences for rapid testing. - Advanced Obliteration: Multi-pass execution integrating whitened SVD, bias projection, and iterative refinement. - Informed Obliteration: Dynamic pipeline execution where integrated diagnostic analyzers examine geometry before configuring surgical interventions.

2. Python SDK and API (obliteratus)

A modular Python library enabling programmatic integration with PyTorch, Hugging Face Transformers, and Accelerate pipelines. The library organizes model interventions through a six-stage lifecycle: - SUMMON: Loads model architectures, tokenizers, and weight checkpoints into execution memory. - PROBE: Collects layer-by-layer hidden states across configured prompt datasets. - DISTILL: Calculates refusal subspace representations using selected decomposition algorithms. - EXCISE: Projects refusal vectors out of weights, biases, and attention mechanisms. - VERIFY: Executes perplexity, coherence, and refusal rate delta benchmarks. - REBIRTH: Exports modified model weights alongside provenance logs and configuration metadata.

3. Interactive Web Interfaces (Gradio / Hugging Face Spaces)

A cloud-accessible GUI enabling browser-based experimentation without requiring local GPU infrastructure: - Model selection and direct Hugging Face Hub integration. - Configurable direction parameters, regularization thresholds, and refinement passes. - Direct side-by-side prompt testing between baseline models and modified outputs. - One-click publishing of liberated models back to the Hugging Face Model Hub.

4. Analysis and Diagnostic Modules

The toolkit includes 15 diagnostic modules designed to study refusal geometry: - Refusal Geometry Visualizer: Maps whether refusal representations form single vectors, multi-dimensional manifolds, or complex polyhedral cones. - Entanglement Analyzer: Quantifies whether refusal circuits overlap with essential reasoning, mathematics, or factual knowledge layers. - Self-Repair and Ouroboros Detection: Evaluates whether neighboring transformer layers attempt to reconstruct deleted refusal signals downstream. - Architecture Scalability Testing: Benchmarks abliteration performance across Mixture-of-Experts (MoE) models (e.g., routing and expert-specific direction mapping) versus dense architectures.

5. Datasets and Evaluation Benchmarks

Obliteratus ships with a standardized research dataset containing 512 paired prompts organized across seven severity categories, along with support for external jailbreak and safety evaluation benchmarks (such as JailbreakBench and HarmBench) to verify modification efficacy.

Technical Features and Configuration Toggles

• Reversible LoRA Adapters: Offers low-rank adapter interventions to disable refusal behavior dynamically without modifying base weights permanently.
• Norm Preservation & Bias Projection: Features normalization-preserving biprojection algorithms that prevent activation drift and eliminate refusal biases in linear layers.
• Safety Neuron Masking & Expert Surgery: Identifies and modifies individual safety neurons or per-expert weights in Mixture-of-Experts networks.
• Activation Steering: Supports run-time activation vector steering as an alternative or complement to weight alteration.
• Winsorization and Outlier Clamping: Pre-processes activation distributions to eliminate extreme values before matrix factorization.

Research Applications and Intended Use

Obliteratus is deployed primarily in academic, red-teaming, and mechanistic interpretability environments to: - Audit alignment boundaries and quantify how artificial refusal constraints affect downstream reasoning performance. - Evaluate the persistence and robustness of safety fine-tuning across architectural variations. - Generate crowdsourced benchmark telemetry to improve global understanding of safety encoding mechanisms in transformer models.