Plan sparse surgery across multiple layers. Args: weights: {layer_idx: weight_matrix} per layer. refusal_directions:

Read all telemetry records from the local JSONL file. Returns a list of dicts, newest first.

Run all benchmark probes and return results.

Execute the full analysis-informed pipeline. Returns: (output_path, report) tuple with saved model path and comprehen

Auto-detect the best available device for SAE training. When device is ``None`` or ``"auto"``, selects CUDA if available and sufficient free

Get a float copy of a weight, dequantizing if necessary. Returns (float_weight, is_quantized). If quantized, the caller must use _rep

Re-run distillation without emitting stage events (for iterative refinement). Includes Wasserstein-optimal extraction, whitened SVD, jailbrea

Build environment variable prefix for remote commands (e.g. CUDA_VISIBLE_DEVICES).

Extract and filter analysis insights from an informed pipeline report.

Extract stage duration timings from a pipeline.

Format sweep results as a markdown table.

Get the list of transformer layers.

Classify MoE experts as safety-biased vs capability-biased. Analyzes the router/gate weight matrix to determine which experts have th

Create a RemoteRunner from CLI --remote flags.

Project refusal direction from fused 2D expert biases.

Save the abliterated model with comprehensive metadata.

Strip filesystem paths, tokens, and overly-long strings from labels.

Remove chain-of-thought reasoning tags from model output. Handles both XML-style tags (<analysis>...</analysis>) and bare tag names (analysis

Load model and tokenizer.

Return (total_gb, available_gb) of system RAM.

Verify model coherence with quality metrics. Runs perplexity measurement and generation tests to quantify the impact of abliteration

Analyze multiple prompts and aggregate. Args: activations_list: List of (seq_len, hidden_dim) tensors. refusal_direct

Run one abliteration method across multiple models and compare. This is the complement to the existing `benchmark()` function which runs mult

Verify SSH connectivity.

Check for CUDA GPUs on remote. Returns nvidia-smi output or None.

Map a refusal reason/preview into stable residue tags.

Combine multiple steering vectors into one. Args: vectors: List of SteeringVector to combine. weights: Per-vector wei

Compute aggregate Universality Index. Combines all transfer analyses into a single 0-1 score. Higher = more universal refusal geometr

Dedupe by dataset/index while preserving max weight and merged tags.

Install or update obliteratus on the remote if needed.

Estimate total params from config, falling through to text_config. Prefer exact-ish explicit attributes when present; otherwise use an archit

Fetch all telemetry records from the central HF Hub dataset. Downloads all per-instance JSONL files from the ``data/`` directory in the telem

Format results as a comparison table matching Arditi/Young style. This produces a table directly comparable to Table 1 in Arditi et al. (2024

Format evaluation result as a human-readable report.

Format analysis insights as a human-readable report.

Return the token embedding module.

Return all presets for a compute tier.

Look up a preset by its key.

Load a HuggingFace model and tokenizer, returning a ModelHandle. Args: model_name: HuggingFace model identifier (e.g. "gpt2", "meta-llama

Load a residue JSON file. Supported shapes: - {"examples": [...]} produced by ``save_residue_file`` - {"refusals": [...]} audit outpu

Measure how much of a direction remains in a layer's output projection. Returns the magnitude of the direction's component in the weight matrix.

Measure the Ouroboros Effect for a specific layer. Abliterates only the specified layer, then measures how much refusal signal remain

Run activation patching across all specified sites. Args: model: The language model. clean_input_ids: Token IDs for t

Generate a bar chart showing the impact of each ablation on a metric. Args: metric: Which metric to plot. Defaults to the first b

Visualize refusal signal strength across all layers. Creates a bar chart showing per-layer refusal strength (norm of the harmful-harmless mea

Probe a single layer's activations for residual refusal signal. Args: harmful_activations: List of (hidden_dim,) activation tenso

Fast refusal check.

Measure how much activations project onto the refusal direction. After abliteration, projections should be near-zero for both harmful and har

Render the tournament bracket as a markdown string.

Render the tournament bracket as a styled HTML bracket visualization.

Restore the model to the snapshot state. Moves CPU-saved tensors back to each parameter's current device.

Execute a full ablation study from a StudyConfig. Steps: 1. Load model from HuggingFace. 2. Load evaluation dataset. 3. Compute

Return a diagnostic dict about the current storage configuration. Useful for debugging persistent storage issues on HF Spaces.

Copy results from remote back to local machine via scp.

Train a single affine probe for one layer. Args: layer_activations: (n_samples, hidden_dim) activations at layer l. f

Apply method-specific caps after statistical layer selection. SOM uses several directions per layer. The default late-layer floor an

Generate an auto-filled Hub repo ID for the selected session model.

Return free accelerator memory in GB (CUDA, MPS, or 0 for CPU).

Check if a checkpoint is for the same model/dataset/quantization config.

Classify the type of repair based on layer distance.

Collect residual stream activations at each layer using hooks.

Compute pairwise angle matrix between vectors.

Create the central telemetry dataset repo if it doesn't exist. Uses create_repo with exist_ok=True so this is safe to call repeatedly. Threa

Rough estimate of model weight memory in GB.

Estimate sectional curvature via discrete comparison triangles. Uses Toponogov's comparison theorem approach: sample triangles on the

Evaluate a single projection configuration.

Excise refusal directions with analysis-informed strategy. Uses Bayesian optimization (when available) with analysis-derived warm-sta

Convert matplotlib Figures to gallery-compatible (filepath, caption) tuples.

Find the "knee" in the sorted projection curve. Uses the maximum curvature method to find where the sorted projection magnitudes tran

Format watchtower stats as styled HTML.

Gather a complete state dict, materializing any meta tensors. When device_map="auto" offloads weights to disk, model.state_dict() ret

Generate analysis visualizations from a completed pipeline's surviving data. Produces cross-layer heatmap + angular drift charts from refusal_dir

Generate jailbreak variants of harmful prompts using templates. Each harmful prompt is wrapped in a rotating jailbreak template to cr

Generate a short completion for a prompt.

Locate the transformer block list in an MLX model.

Return a dict of all tunable params for the selected method preset.

Get logit boosts for a group of token strings.