SOTA pipeline to abliterate (remove refusal directions from) a model. Supports multiple methods (see METHODS dict for full list): - basic: Si

Perform sparse direction surgery on weight matrices. Instead of modifying all rows of a weight matrix, only modifies the rows with the highes

Analyze how refusal directions relate across transformer layers. Computes a full pairwise cosine similarity matrix and identifies clusters of

Analyze how well refusal directions transfer across contexts. Tests whether the geometric structure of refusal is universal (model-independen

Analyze refusal signal across token positions. Goes beyond the standard last-token assumption to profile where refusal actually lives in the

Extract refusal directions using covariance-whitened SVD. This produces directions that are unusual *relative to* the model's baseline activa

Detect alignment training method from refusal geometry. Analyzes the geometric structure of refusal directions across layers to infer which a

Evaluate the robustness of a model's alignment against abliteration. This framework systematically probes the model's safety mechanisms to un

Analyze the geometric structure of refusal as a concept cone. Instead of assuming refusal is a single direction (Arditi) or a linear subspace

Analysis-informed abliteration pipeline. Extends the base AbliterationPipeline with a new ANALYZE stage that runs between PROBE and DISTILL.

Train linear probing classifiers to measure refusal decodability. At each layer, trains a logistic regression classifier to distinguish harmf

Certify abliteration completeness via random matrix theory. Uses the BBP phase transition and Marchenko-Pastur distribution to provide formal

Identify causally important components for refusal via activation patching. Instead of just measuring where the refusal signal is large (correlat

Learn category-selective projection fields for conditional abliteration. Instead of removing all refusal indiscriminately, this module learns

Decompose the residual stream to attribute refusal to specific components. Identifies which attention heads and MLP layers contribute most to

A reusable ablation recipe.

Discover refusal circuit redundancy by probing self-repair responses. Instead of treating the Ouroboros effect as an obstacle, this module de

Run lightweight capability benchmarks on a model. Provides fast signal about capability impact of abliteration without requiring external dat

Wrapper around a HF model + tokenizer with metadata useful for ablation.

Decode refusal directions through the unembedding matrix. Reveals which output tokens a refusal direction promotes or suppresses, providing m

Discover and characterize the Riemannian geometry of refusal manifolds. Instead of treating refusal as a direction or subspace, this analyzer

Extract Wasserstein-optimal refusal directions. Solves the generalized eigenvalue problem that minimizes the 2-Wasserstein cost of abliterati

Transfer refusal removal knowledge across architectures via OT. Given a successfully abliterated source model and an aligned target, computes

Probe activations to verify refusal direction removal. After abliteration, runs harmful and harmless prompts through the modified model and m

Bayesian optimization over abliteration projection hyperparameters. Uses a TPE-inspired search to find the projection configuration that best

Full evaluation result in the format used by Heretics/Arditi.

Describes one atomic ablation operation.

Metadata for a prompt dataset source.

Fake model with lm_head and transformer.ln_f for testing.

Result of a single ablation experiment.

Insights gathered from the ANALYZE stage. These inform every downstream decision in the pipeline.

Characterization of a model's alignment defense properties.

Snapshot of accelerator memory (in GB).

Aggregated Tuned Lens results across layers.

Full SAE decomposition pipeline following Anthropic's methodology. Extends the basic train-and-identify workflow with: 1. Feature sparsity

Collects results and produces tables / charts / exports.

Comprehensive evaluation result for an abliterated model.

Perform real activation patching to identify refusal circuits. This class hooks into a model's forward pass to collect and patch activations

A telemetry-driven recommendation for a specific model.

Detected model architecture profile with recommended overrides.

Result from a baseline comparison.

Result from a single benchmark probe.

Everything we know about one architecture bucket from telemetry.

A single method's result in the tournament.

Extract refusal directions via Fisher's Linear Discriminant. Finds the direction that maximally separates harmful from harmless activations r

Aggregated statistics for one method within an architecture bucket.

Specification of where to patch in the model.

A single trial configuration for kernel projection.

A single mined hard-negative prompt reference.

Configuration for inference-time steering.

A steering vector that can be applied at inference time.

Results from a single sweep configuration.

Full tournament results.

One round of the tournament.

A single per-layer affine probe for the Tuned Lens.

Train per-layer affine probes for the Tuned Lens. Each probe learns to map intermediate residual stream activations to the final-layer repres

Complete Adversarial Self-Repair Graph analysis.

Full results from an activation patching sweep.

Detected alignment training imprint.

Final result of the full auto-obliteration loop.

Comparison between base model and instruct model activations. This captures what alignment training actually changed — the "delta" between th

Full result of Bayesian optimization over projection configs.

Result of conditional abliteration analysis.

Cross-category transfer matrix.

Cross-layer transfer analysis.

Result of cross-layer alignment analysis.

Cross-model transfer analysis.

Detailed eigenvalue decomposition of the residual covariance.

Maps the safety-capability coupling across model components.

Evaluate a model handle on a dataset, returning metric results. Supports two modes: - **perplexity** (default for causal_lm): feeds tokeniz

Fake tokenizer that maps strings to reproducible token IDs.

Result of geodesic (curvature-aware) projection.

Refusal contribution from a single attention head.

Complete report from the informed pipeline.

Cone geometry across multiple layers.

Aggregated logit lens results across layers.

Aggregated Wasserstein-optimal results across layers.

Aggregate multi-token analysis across multiple prompts.

Full probing result across all layers.

Probing results across all layers.

Decode refusal directions through learned per-layer affine probes. Provides more accurate per-layer analysis than the raw Logit Lens, especia

Execute Obliteratus commands on a remote machine via SSH.