Test whether the extracted refusal direction is specific to harmful prompts. Computes the ratio of harmful-to-harmless projection magnitudes.

Release cached allocations on the current accelerator.

Package all research artifacts from the last obliteration into a downloadable archive. Exports: - refusal_directions.pt: Per-layer refusal di

Return a list of ModelPreset objects for open-weight models in the BESTIARY catalog that aren't already curated. Empty list on any failure (safe n

Extract LEACE directions for all layers. Args: harmful_acts: {layer_idx: [activations]} from activation collection. h

Format full decomposition pipeline results.

Format activation patching results as a report.

Format probe results as a human-readable report.

Format Bayesian optimization results.

Format Wasserstein-optimal extraction results.

Format single-layer cone analysis as a report.

Format cross-layer analysis as a human-readable report.

Format multi-layer logit lens analysis as a report.

Forward pass returning reconstruction and latent features.

Aggressive memory cleanup: GC + accelerator cache flush.

Generate text using the MLX model. Parameters ---------- handle : MLXModelHandle A loaded MLX model handle. prompt : str

Extract hidden-state activations from specified layers. Uses MLX's computation graph to capture intermediate outputs. Parameters -------

Return all tracked models.

Query memory for the given accelerator (or system RAM for MPS/CPU).

Get a specific model by ID.

Return models with status 'new' (not yet queued or obliterated).

Sum of free memory across all accelerator devices, in GB.

Retrieve a weight matrix from the model. Parameters ---------- handle : MLXModelHandle Loaded model. layer_idx : int

Default AnalysisInsights for testing.

True if *any* GPU backend (CUDA or MPS) is available.

Return True if *exc* is an out-of-memory error on any backend.

True if the scheduler is actively running.

Convenience: yield specs one at a time, applying + restoring around each.

Re-run abliteration with a benchmark config and load result into Chat. On ZeroGPU, uses the visitor's GPU quota.

Load a HuggingFace model via ``mlx-lm`` for Apple-native execution. Parameters ---------- model_name : str HuggingFace model iden

Convert an MLX array to a PyTorch tensor.

A minimal mock transformer model. Provides: - model.config with config.num_hidden_layers = 4 - model.named_parameters() returning fake w

A minimal mock tokenizer with encode, decode, and apply_chat_template.

Just the ids — the common case for populating a dropdown or a config.

Modify a weight matrix in-place using a function. Parameters ---------- handle : MLXModelHandle Loaded model. layer_idx : int

Multi-objective: minimize (refusal_rate, kl_divergence).

Run the full obliteration pipeline, streaming log updates to the UI. On ZeroGPU Spaces, this function runs on the visitor's GPU quota (up to

An InformedAbliterationPipeline with no model loaded.

Plot the capability vs safety tradeoff.

Visualize the logit lens token promotion/suppression spectrum.

Project a direction out of a weight matrix (abliteration). Given weight matrix W and unit direction d, computes:: W' = W - (W @ d) outer

Push a session model to HuggingFace Hub, with optional refinement.

Ablate random directions as a negative control. For each trial: 1. Generate a random unit vector in activation space 2. Project it out of

All methods ranked by mean score, descending.

A normalized random torch tensor of shape (768,).

Class decorator that registers a strategy under its `name` attribute.

Clear all state and start fresh.

Run a hyperparameter sweep over abliteration configurations. For each combination of sweep_params (crossed with n_seeds random seeds), runs t

Run an elimination tournament across selected abliteration methods. Each individual method is run inside its own ``@spaces.GPU`` allocation (

Return a dtype safe for SVD on the tensor's device. MPS does not support float64, so we cap at float32.

Save the (modified) MLX model to disk. Saves in safetensors format compatible with both MLX and HuggingFace. Parameters ---------- h

Set random seed on all available accelerators.