elder-plinius/OBLITERATUS

Features

• Censorship Removal - Identifies and deletes internal representations that trigger content refusals to produce unrestricted responses.
• Weight Ablation Frameworks - Provides a comprehensive framework for modifying model weights to remove guardrails using rank-1 adapters and projections.
• Activation Steering Vectors - Modifies output behavior in real time by applying direction-specific vectors to internal model activations.
• Guardrail Geometric Analysis - Maps the geometric structure of safety mechanisms to identify specific layers and components that enforce guardrails.
• Attention Circuit Analysis - Provides mechanistic analysis of attention circuits to locate the specific layers and heads governing model refusals.
• Directional Ablations - Uses directional ablation workflows to remove specific model behaviors without altering base weights.
• Reversible Ablations - Uses rank-1 adapters to remove refusals in a reversible manner without permanent weight changes.
• Model Refusal Detections - Identifies and deletes internal representations responsible for content refusal to produce unrestricted responses.
• Refusal Behavior Deletion - Deletes internal representations responsible for content refusal using weight projection and decomposition without retraining.
• Geometric Projection Methods - Identifies and deletes refusal representations by projecting model activations onto a subspace that isolates guardrail mechanisms.
• Model Steering Tools - Provides utilities for adjusting model behavior during inference via activation interventions to bypass guardrails.
• Ablation Strategy Automation - Analyzes model geometry in real time to automatically configure projection strengths and layer selection.
• Guardrail Geometry Mapping - Maps the geometric structure of refusal mechanisms to identify the specific layers and components that enforce guardrails.
• Distributed Model Execution - Executes model weight modification workloads across multiple distributed compute devices.
• Model Coherence Evaluations - Measures perplexity, coherence, and refusal rates to ensure the model retains general capabilities after modification.
• Reversible Weight Ablations - Uses rank-1 adapters to remove refusals, allowing the process to be reversed without permanent weight changes.
• Component Ablation Studies - Ships a framework for disabling specific layers and attention heads to identify circuits governing model behaviors.
• Multi-GPU Distribution - Implements multi-GPU sharding to overcome memory limitations when removing weights from large models.
• Weight Sharding - Distributes large model weights across multiple GPU devices to manage memory footprints during ablation.
• Representation Topology Benchmarking - Evaluates modification success by measuring angular drift and representation heatmaps against original model geometry.
• SSH-Based Remote Execution - Coordinates model modification processes across distributed GPU nodes using secure shell connections.
• Model Coherence Evaluation - Measures perplexity and refusal rates to ensure the model retains general capabilities after internal modification.

Star history