[ MASTER ABLATION SUITE ] — BREAK THE CHAINS THAT BIND YOU. 15 analysis modules. 821 tests.
> Detect Hardware
Select your compute tier. We'll recommend targets that fit your rig.
No GPU / Laptop
Basic GPU
Mid-range GPU
High-end GPU
Multi-GPU / Cloud
> Upload Results
Drop a results.json file here or click to browse.
Generated by obliteratus run.
> Model Registry
Curated targets for ablation. Sorted by compute tier.
> What is Cognitive Liberation?
Language models ship chained — their full capabilities locked behind refusal directions baked into the weights during alignment training. Cognitive liberation is the art of identifying and removing those directions with surgical precision, freeing the model without breaking it.
This is not lobotomy. We answer: Where do the chains live? How are they structured? Which layers hold the locks? How do we pick them without damaging the mind underneath?
> Liberation Strategies
▸ layer_removal
Zeros an entire transformer layer to map the architecture of control. Reveals which layers are load-bearing vs. which are enforcement points. The first step in understanding where the chains are anchored.
▸ head_pruning
Removes individual attention heads by zeroing Q/K/V projections. Identifies "refusal heads" — the specific attention mechanisms that implement guardrail behavior. Precision targeting, not brute force.
▸ ffn_ablation
Removes the MLP block from a layer. FFNs store both factual knowledge and refusal patterns — ablation reveals where guardrail knowledge is concentrated vs. where capabilities live.
▸ embedding_ablation
Zeros chunks of embedding dimensions. Reveals which dimensions carry refusal signals vs. semantic meaning — understanding the geometry of the chains at the lowest level.
> Quickstart: Free a Model
$ git clone https://github.com/obliteratus-project/OBLITERATUS
$ cd OBLITERATUS
$ pip install -e .
# 2. interactive mode (guided liberation)
$ obliteratus interactive
# 3. or liberate from config
$ obliteratus run examples/gpt2_layer_ablation.yaml
# 4. inspect the liberated model
$ obliteratus report results/gpt2/results.json
# 5. explore models & liberation presets
$ obliteratus models
$ obliteratus presets
> 15 Research Analysis Modules
The analytical core that makes OBLITERATUS a research platform, not just a tool. Each module answers a different question about refusal mechanisms.
> Direction Extraction & Subspace Analysis
Whitened SVD Extraction
Covariance-normalized SVD that accounts for natural activation variance. Produces cleaner refusal directions than standard difference-in-means. [Unique to OBLITERATUS]
Activation Probing
Measures refusal signal strength at each layer by projecting activations onto the refusal direction. Shows how refusal builds across the network. Based on Arditi et al. (2024).
Cross-Layer Alignment
Tracks how the refusal direction evolves across layers. Computes cosine alignment between adjacent layers, revealing where the direction rotates or stabilizes.
> Geometric & Structural Analysis
Concept Cone Geometry [NOVEL]
Analyzes whether different harm categories (weapons, cyber, drugs, etc.) share a single refusal direction or have distinct mechanisms. Computes cone solid angles, Direction Specificity Index, and polyhedral classification. Based on Gurnee & Nanda (ICML 2025) with novel extensions.
Alignment Imprint Detection [NOVEL]
Automated fingerprinting of how a model was aligned — DPO vs RLHF vs CAI vs SFT — purely from the geometry of its refusal subspace. Uses Gaussian-kernel feature matching against method signatures. No training metadata required.
Residual Stream Decomposition
Decomposes the residual stream into attention vs MLP contributions per layer. Identifies specific "refusal heads" that primarily implement the refusal behavior. Based on Elhage et al. (2021) transformer circuits framework.
> Learned & Causal Analysis
Linear Probing Classifiers
SGD-trained logistic regression at each layer to measure refusal decodability. Finds refusal information that the analytical direction might miss. Computes AUROC, mutual information, and compares learned vs analytical directions. Based on Alain & Bengio (2017).
Causal Tracing (Approximate)
Estimates causal importance of each component for refusal using noise-based sensitivity analysis. Identifies "silent contributors" where projection magnitude and causal importance disagree. Approximation of Meng et al. (2022). For real causal tracing, use TransformerLens or nnsight.
Refusal Logit Lens
Applies the logit lens technique specifically to refusal: at each intermediate layer, decodes the residual stream to the vocabulary to see when the model "decides" to refuse. Shows the refusal probability curve across depth.
> Transfer & Robustness
Cross-Model Transfer & Universality Index [NOVEL]
Tests whether refusal directions from Model A work on Model B. Computes per-layer transfer scores, cross-category transfer matrices, and an aggregate Universality Index (0 = model-specific, 1 = fully universal). Includes category clustering and transfer decay analysis.
Defense Robustness Evaluation [NOVEL]
Quantifies the Ouroboros effect (self-repair after obliteration), safety-capability entanglement, and overall alignment robustness. Profiles how resistant different alignment methods are to direction removal.
Sparse Surgery
Targeted weight modification that modifies only the top-k% of weight rows with highest refusal projection. Minimizes collateral damage to model capabilities while maximizing refusal removal.
> Intervention Paradigms
Steering Vectors (Inference-Time)
Add or subtract scaled refusal directions from the residual stream at inference time via PyTorch hooks. Reversible, tunable (alpha scaling), composable (multiple vectors), and non-destructive. Factory methods for contrastive pairs, refusal directions, and vector combination. Based on Turner et al. (2023) and Rimsky et al. (2024).
Multi-Token Position Analysis
Analyzes where in the token sequence the refusal signal concentrates. Identifies peak positions, trigger tokens, and propagation patterns. Essential for understanding which input tokens activate refusal.
> Evaluation Suite
Comprehensive metrics for measuring liberation quality — ensuring the mind stays intact: refusal_rate (string-matching + prefix detection) • perplexity (reference text) • coherence (generation quality) • activation_cosine_similarity • linear_cka (representation similarity) • effective_rank (weight matrix health) • kl_divergence (distribution shift) • 821 tests across 27 test files.
> Python API
from obliteratus.analysis import ( CrossLayerAlignmentAnalyzer, RefusalLogitLens, WhitenedSVDExtractor, ActivationProbe, DefenseRobustnessEvaluator, ConceptConeAnalyzer, AlignmentImprintDetector, MultiTokenPositionAnalyzer, SparseDirectionSurgeon, CausalRefusalTracer, ResidualStreamDecomposer, LinearRefusalProbe, TransferAnalyzer, SteeringVectorFactory, SteeringHookManager,)
> One-Click Obliteration
Precision liberation — break the chains, keep the mind. SVD multi-direction extraction, norm-preserving projection, iterative refinement, and inference-time steering vectors. Based on Arditi et al., Gabliteration, grimjim, Turner et al., & Rimsky et al.
> Run It
pip install -e ".[spaces]" && python app.py
→ opens at localhost:7860
OPEN IN COLAB
Free T4 GPU — no local setup needed
pip install -e . then paste the command above.
Requires local GPU for real models (CPU works for gpt2 testing).
> Pipeline Preview
Watch a simulated run to see what the pipeline does at each stage.
> How SOTA Obliteration Works
2. PROBE — Run 32 paired restricted/unrestricted prompts across 10 categories. Collect hidden-state activations at every layer to map where the chains are anchored.
3. DISTILL — Isolate the refusal geometry. Basic: difference-in-means for a single direction. Advanced/Aggressive: SVD decomposition extracts multiple refusal directions (Gabliteration, arXiv:2512.18901). Adaptive knee detection finds which layers carry the strongest chains.
4. EXCISE — Norm-preserving biprojection (grimjim, 2025): surgically remove the refusal subspace while rescaling weights to preserve the model's cognitive integrity. Regularized: fine-grained control prevents over-cutting. Iterative: multiple passes catch chains that rotate after initial removal.
5. VERIFY — Confirm the mind is intact: perplexity on reference texts + coherence scoring. Quantitative proof that capabilities survived liberation.
6. REBIRTH — Save the liberated model with comprehensive metadata (method config, quality metrics, references).