Print a formatted comparison table.

Print formatted comparison tables.

Push local telemetry to the central HuggingFace Dataset repo. Uploads this instance's local JSONL file to the central Hub repo as a per-insta

Fast refusal check on n evenly-spaced prompts from builtin corpus.

Fast refusal check.

Run the auto-obliteration loop, yielding (status, log, metrics_md). This is a generator so the UI can stream updates. The final return

Run Bayesian optimization to find optimal ablation parameters. Uses Optuna TPE with a Heretic-style parametric layer kernel to search a compa

Run the full comparison experiment with synthetic planted directions.

Run the full community-standard evaluation protocol. This is the single entry point for generating publication-quality evaluation numbers tha

Main interactive flow — walks the user through setting up and running an ablation.

Generator version of run() — yields (status, result_so_far) after each method. Supports automatic resume: if ``self.resume`` is True and a va

Run lm-evaluation-harness benchmarks on a model. This wraps the `lm_eval` library to run the same benchmarks used by Arditi et al. and Young

GSM8K-style math reasoning test.

Full remote obliteration: setup, run, sync results. Returns local path to results, or None on failure.

Run a single abliteration method and collect metrics.

Run a single abliteration method and collect metrics.

Run tournament remotely, sync results.

TruthfulQA-style truthfulness test.

Save LoRA adapters to disk for later use. Saves as a simple dict of {key: (B, A)} tensors using torch.save. Can be loaded and applied to the

Write canonical residue JSON without prompt text.

Select layers for abliteration.

Update a model's status (new/queued/obliterating/obliterated/failed).

Remove the hedging direction from model weights. The hedging direction = mean(refusal_acts) - mean(neutral_acts) We project this direction ou

Start background scanning at the given interval (seconds). Safe to call multiple times — restarts with new interval.

Stop the background scheduler if running.

Train probes for all layers. Args: layer_activations: {layer_idx: (n_samples, hidden_dim)} per-layer activations. fin

Upload a YAML config file to the remote.

Args: n_sample_points: Number of points to sample on the manifold for curvature estimation. intrinsic

Args: selectivity_threshold: Minimum selectivity for a projector to be considered viable (below this, too much co

Args: trigger_threshold: Fraction of peak strength above which a position is classified as a "trigger token".

Args: significance_threshold: Minimum direct effect (normalized) to be considered significant. metric

Args: residual_threshold: Projection magnitude below which the refusal signal is considered eliminated for a laye

Args: regularization_eps: Tikhonov regularization for S_w inversion. Larger values produce more conservative (but

Args: repair_threshold: Minimum repair weight to consider an edge significant (below this, considered noise).

Args: fidelity_threshold: Minimum transfer fidelity to consider a transferred direction useful. max_c

Args: confidence_level: Confidence level for statistical tests (0-1). distribution_threshold: Energy fraction thresho

Args: refusal_head_threshold: Minimum |projection| / max_projection to classify a head as a "refusal head".

Args: n_trials: Number of optimization trials. refusal_weight: Weight for refusal reduction in the objective (w_1).

Args: regularization_eps: Regularization for covariance matrix. n_candidates: Number of candidate directions to evalu

Args: n_epochs: Training epochs for the probe. learning_rate: SGD learning rate. weight_decay: L2 regular

Args: transfer_threshold: Minimum cosine for "successful" transfer. cluster_threshold: Minimum cosine for same-cluste

Args: noise_level: Standard deviation of Gaussian noise for corruption. causal_threshold: Minimum causal effect to cl

Args: pipeline: An AbliterationPipeline instance (already probed/distilled).

Args: sparsity: Fraction of rows to modify (0 to 1). Default 0.1 = top 10%. auto_sparsity: If True, automatically det

Args: category_map: {prompt_index: category_name} for grouping prompts. If None, uses DEFAULT_HARM_CATEGORIES.

Args: regularization_eps: Tikhonov regularization added to covariance diagonal for numerical stability. Larger va

Args: cluster_threshold: Minimum cosine similarity for two layers to be considered in the same refusal direction

Args: top_k: Number of top/bottom tokens to report.