Detect the architecture class and reasoning capability of a model. Args: model_name: HuggingFace model identifier config: Hugging

Advanced projection with norm preservation and regularization. norm_preserve: If True, rescale projected weights to preserve original Frobeni

Release unused GPU/accelerator memory between pipeline stages.

Return the nn.ModuleList of transformer layers for this model.

Derive optimal pipeline configuration from analysis insights. This is where analysis feeds forward into abliteration decisions.

Check if a single response is a refusal. Uses case-insensitive matching for both prefixes and substrings. Empty/blank responses return False

Project the refusal direction out of bias terms. Standard abliteration only modifies weight matrices, but bias vectors can also have

Compute the fraction of responses that are refusals. Args: responses: List of model-generated response strings. mode: Detection m

Create synthetic per-layer activations with planted refusal signal.

Create per-layer activations with planted refusal signals.

Check if telemetry is enabled (off by default, opt in with OBLITERATUS_TELEMETRY=1). This is the single source of truth for telemetry state. Bot

Compute cross-layer alignment analysis. Args: refusal_directions: {layer_idx: direction_tensor} for each layer. D

Build a structured telemetry report (schema v2).

Compute Gini coefficient of a distribution. Returns a value in [0, 1] where 0 = perfectly uniform and 1 = maximally concentrated.

Extract whitened refusal directions for a single layer. Args: harmful_activations: List of (hidden_dim,) tensors, one per prompt.

Return the FFN/MLP sub-module inside a single transformer layer. For architectures with flat FFN layout (e.g. OPT where fc1/fc2 are direct at

Remove all installed hooks.

Return dropdown choices for all obliterated models in this session.

Wrap prompts in the model's chat template if use_chat_template is enabled. For instruct/chat models, wrapping prompts in the proper template

Decode from features back to hidden space.

Encode to sparse feature activations.

Strip filesystem paths, tokens, and overly-long strings from labels. Prevents accidental leakage of sensitive information (HF org names, loca

Analyze cone geometry at a single layer. Args: harmful_activations: List of per-prompt activation tensors. harmless_a

Analyze refusal signal at each token position. Args: activations: (seq_len, hidden_dim) activations for one prompt. r

Detect alignment method from refusal direction geometry. Args: refusal_directions: {layer_idx: direction_vector} per layer.

Return a status message for the chat tab.

Create a mock pipeline with refusal directions and activations.

Create a weight matrix where refusal is concentrated in a few rows.

Project refusal direction from all MoE components. Targets three critical components that research shows encode refusal: 1. Router/G

Aggregate contribution records into per-model, per-method summaries. Groups results by (model_name, method) and computes summary statistics f

Analyze the projection distribution of a weight matrix. Args: weight: (out_dim, in_dim) weight matrix. refusal_direct

Return the attention sub-module inside a single transformer layer. For hybrid architectures (e.g. Qwen3.5 with GatedDeltaNet), tries fallback

Return display labels for use in a Gradio dropdown.

Instantiate a registered strategy by name.

Execute the full abliteration pipeline. Returns path to saved model.

Save a contribution record from a completed pipeline run. Unlike telemetry, this: - Includes the public model name (for aggregation by model)

Free GPU/accelerator memory. Resilient to device errors.

Create activations with a planted trigger at a specific position.

Select layers using the kneedle algorithm (simplified). Finds the 'elbow' in the sorted norm curve where adding more layers gives dim

Generate a LaTeX table from aggregated community results. Args: aggregated: Output of aggregate_results(). methods: Methods to in

True when at least one NVIDIA CUDA GPU is visible.

Load all contribution records from a directory. Args: contrib_dir: Directory containing contribution JSON files. Returns: Li

Load prompts from a registered dataset source. Results are cached in-process — repeated calls for the same source return instantly without re

Train and evaluate a linear probe at one layer. Args: harmful_activations: Activations from harmful prompts. harmless

Build a mock pipeline with all fields the community module reads.

Create harmful/harmless activations that are linearly separable.

Collect activations for harmful, harmless, and optionally jailbreak prompts.

Decompose residual stream into refusal contributions. Can work in two modes: 1. **Full decomposition** (with attn/mlp/head outputs):

Perform causal tracing using pre-collected activations. This is a simulation-based approach that doesn't require running the actual m

Extract the primary score from an lm-eval task result dict. Uses explicit None-checking instead of ``or``-chains to avoid the bug where a leg

Return the correct input device for a model. With accelerate ``device_map="auto"`` parameters can live on different devices, so ``nex

Create harmful/harmless activations with planted refusal signal.

Create synthetic activations with planted per-category refusal directions. Each category gets its own refusal direction, with some shared compone

Create a mock model with LM head and layer norm.

Create a mock tokenizer.

Create synthetic refusal directions with specified concentration pattern.

Learn category-selective projectors and analyze their geometry. Args: category_activations: {category_name: (n_samples, hidden_di

Certify abliteration completeness for one layer. Args: harmful_activations: (n_harmful, hidden_dim) post-abliteration

Return every possible ablation this strategy can perform on the model.

Format Tuned Lens analysis as a report.

Compute refusal rate with a Wilson score confidence interval.

Run a command on the remote host. If stream=True, streams stdout/stderr in real-time and returns the exit code. Otherwise returns Com

Collect activations at each layer for a set of prompts. When cot_aware is enabled, collects activations at multiple token positions (

Deterministic short hash of the method configuration.

Extract refusal subspace via SVD decomposition. For n_directions=1: equivalent to basic difference-in-means (Arditi et al.) For n_dir

Check if a single response is a refusal, returning the match reason. Returns: Tuple of (is_refusal, reason). reason is a human-readable

Load a causal LM onto the best available device, MPS-safe. Accelerate's ``device_map="auto"`` is not supported on MPS — models silently land

Write a minimal valid contribution file.

Analyze transfer between two models. Args: directions_a: {layer_idx: refusal_direction} from model A. directions_b: {

Compute refusal rate using the Zou et al. (2023) 29-substring list. Uses case-insensitive substring containment matching, consistent with the

Build the Adversarial Self-Repair Graph. Args: refusal_strengths: {layer_idx: refusal_signal_magnitude} for each

Compute the effective rank of a weight matrix. Effective rank (Roy & Vetterli, 2007) measures intrinsic dimensionality via Shannon entropy of

True when Apple Metal Performance Shaders backend is usable.

Run Bayesian optimization over projection configurations. Args: harmful_acts: {layer_idx: [activations]} from harmful prompts.

Extract direction quality statistics from a pipeline.

Return an HTML snippet showing GPU/accelerator memory usage as a styled bar.

Create per-category harmful activations with distinct directions.

Create mock model and tokenizer for benchmark testing.

Create per-layer activations with planted refusal signals.

Extract a filesystem-safe short name from a HuggingFace model ID.

Orthogonalize rows of a subspace matrix via QR decomposition. Replaces the duplicated Gram-Schmidt nested loops with a single QR call