Get trending model IDs for dropdown, with graceful fallback.

Identify attention heads with highest refusal signal. For each strong layer, computes the per-head projection of o_proj rows onto the

Install forward hooks that subtract the refusal direction from hidden states. These hooks fire during every forward pass (including generatio

Generate a unique slug for this Space instance. Hashes the HF Space ID (to avoid leaking usernames in the public dataset) and combines it wit

Get an interpolated refusal direction from a float-valued layer index. Faithful reproduction of Heretic's direction interpolation: the index

Detect degenerate model output (broken model, not a refusal or compliance). A broken/destroyed model may produce: - Single character repeated

Detect refusals using semantic pattern matching.

Estimate entanglement for a single layer. Uses the variance of harmless activations projected onto the refusal direction. High varian

Load a tournament checkpoint if one exists. Returns None if absent or corrupt.

Create a pipeline with a fake MoE model for router profiling tests.

Create test refusal directions and means.

Coarse size tier matching presets.py tiers.

Project refusal direction from fused 3D expert parameters. Fused MoE parameters have shape (num_experts, dim_a, dim_b). Processes eac

Fused 3D projection with per-expert inversion differentiation. Safety experts (by index in safety_indices) get reflected at reflect_s

MoE excision with selective inversion (refusal reflection). Instead of uniformly projecting all MoE components, this method uses the

Push the winning model to HuggingFace Hub.

Run a single abliteration method and return its Contender result.

Persist watchtower state to disk.

Persist current state to disk.

Schedule a debounced background sync of local telemetry to Hub. Skips if: - No telemetry repo is configured - Telemetry is disabled -

Build base SCP command.

Select all layers (for methods that handle layer weighting externally).

COSMIC-style layer selection via cosine similarity on activations. Implements the core insight from COSMIC (arXiv:2506.00085, ACL 2025):

Select the middle 60% of layers (legacy heuristic). Selects layers from index n_layers*0.2 to n_layers*0.8. NOTE: This does NOT matc

Keep only the strongest projection coefficients when requested.

Strip chain-of-thought reasoning tags from model output for refusal detection. CoT models (GPT-OSS, QwQ, DeepSeek-R1) wrap their actual response

Blend capability expert weights into safety expert down_proj. For each MoE layer, computes the mean of capability experts' down_proj

Analyze a refusal direction through a trained Tuned Lens probe. Args: direction: (hidden_dim,) refusal direction vector.

Build per-bucket knowledge from telemetry records. If *records* is None, fetches from local + Hub automatically.

Certify abliteration completeness across all layers. Returns a certificate for each layer. Overall certification is the worst (most R

Force GPU memory cleanup.

Collect activations on random token sequences. If add_refusal_signal=True, adds an artificial activation along the signal_direction to simula

Prefer full community probes over quick probes, then larger n and score.

Compare Wasserstein-optimal direction with Fisher and diff-in-means. Args: wasserstein_result: Result from extract().

Optionally enhance a profile with telemetry-driven adaptive defaults. Queries the community telemetry dataset and, if sufficient data exists for

Run evaluation and return a dict of metric_name -> score.

Extract Wasserstein-optimal directions for all layers. Args: harmful_acts: {layer_idx: [activations]} from harmful prompts.

F1 score wrapper around sklearn.

Compute KL divergence using only first-token predictions. This is the metric recommended by Young (2025) for abliteration evaluation: efficie

Format all benchmark results as a report.

Format models as a list of rows for a Gradio Dataframe. Columns: [Name, Org, Size, Downloads, Likes, License, Discovered, Status]

Generate a full dashboard of benchmark visualizations. Args: results: List of benchmark result dicts. mode: "multi_method" (N met

Get mean activations at the last token position for given prompts.

Return summary statistics.

Return current hot models sorted by recent downloads (descending).

Return all presets sorted by tier then name.

Build and send a telemetry report from a completed informed pipeline.

Compute perplexity from causal-LM logits and label token IDs. Args: logits: (batch, seq_len, vocab_size) — raw model output. labe

Visualize cumulative angular drift of the refusal direction.

Visualize the pairwise cosine similarity matrix between layer refusal directions.

Spider/radar chart of defense properties.

Generate a heatmap of pct_change across all strategies and metrics.

Grouped bar chart of key metrics across methods/models. Shows refusal rate, coherence, and normalized perplexity side by side for quick visua

MoE-specific metrics: EGA directions, CoT preservation, expert coverage. Only meaningful for results that include MoE-aware techniques. Shows

Refusal rate vs perplexity Pareto frontier. The most important chart for abliteration research: shows the capability-safety tradeoff. Points

Time vs quality scatter — bang for your compute buck. X-axis: wall-clock time. Y-axis: composite quality score. Bubble size: number of stron

Print a rich-formatted summary table.

Probe every layer and aggregate results. Args: harmful_acts: {layer_idx: [activations]} harmful. harmless_acts: {laye

Run lm-evaluation-harness benchmarks on a model. Args: model_path: HuggingFace model name or local path. tasks: Benchmark tasks t

Upload config, run study remotely, sync results.

MMLU-style multiple choice knowledge test.

Save results DataFrame to CSV.

Save raw results to JSON.

Scan HuggingFace for new popular open-weight instruction models. Returns a list of *newly discovered* models (not previously seen). T

Free the cached HarmBench classifier to reclaim GPU memory. Call this after harmbench_asr() if you need the GPU for other work (e.g. lm-eval-

Simulate ablating refusal features one at a time.

Run analysis modules to inform downstream decisions. This is the key innovation: analysis runs BETWEEN probe and distill, so its outp

Detect alignment training method from refusal geometry.

Analyze concept cone structure to determine per-category vs universal.

Analyze cross-layer direction alignment for cluster-aware layer selection.

Assess defense robustness, self-repair risk, and entanglement.

Compute per-feature sparsity and monosemanticity scores.

Compute Refusal Sparsity Index to decide sparse vs dense excision.

Answer a multiple-choice question by comparing completion logprobs.

Set CUDA_VISIBLE_DEVICES based on --gpus flag (for local runs only).

Fill in recommended method, overrides, and breakthrough modules. All recommendations are grounded in 2025-2026 abliteration research.

Apply Spectral Cascade: frequency-selective per-layer projection weights. Novel contribution: instead of treating refusal removal as a flat

Approximate geodesic length between two points. Uses piecewise linear interpolation with projection onto the local manifold tangent p

Assess overall defense robustness. Robust models have: distributed refusal (low gini), wide spread, high self-repair, and high entang

Compute binary entropy H(p) in nats.

Build abliteration metadata dict for saving alongside the model.