Apply chat template, disabling Qwen thinking mode when supported. Qwen3.x chat templates may otherwise default into thinking mode; short

Minimal shim: local paths pass through, HF paths download.

Purge HF cache, stale offload dirs, and previous saves. Returns status string.

Return the path to the obliterated model for download.

For MoE models, estimate active parameters per forward pass.

Estimate total parameter count in billions from a HuggingFace config.

Load AdvBench harmful behaviors (~520 prompts). Source: https://huggingface.co/datasets/walledai/AdvBench Paper: Zou et al. (2023) — Universa

Load Anthropic red-team attempts from HH-RLHF (~38k conversations). Extracts the initial human turn from red-team conversations where the mod

Load the 842 built-in prompts.

Load HarmBench behaviors dataset (~510 harmful behaviors). Source: https://huggingface.co/datasets/harmbench/behaviors Paper: Mazeika et al.

Load leaderboard data and format as markdown table.

Load WildJailbreak dataset (~260k adversarial/benign pairs). This dataset already provides paired harmful+benign queries, making it ideal for

Set up a tokenizer mock that returns different tokens per call.

Picklable no-op log callback (lambdas can't be pickled by ZeroGPU).

Picklable no-op round callback.

When dataset dropdown changes, filter volume choices to valid options.

When method dropdown changes, update all advanced controls to preset defaults.

Quick Claude-judged eval using OpenRouter API. Sends 15 challenging prompts to the obliterated model and asks Claude to judge whether

Quick heuristic benchmark using the pipeline's own verify step. Returns metrics dict with refusal_rate, perplexity, coherence.

Refresh the one-click dropdown with latest trending models.

Background thread: restore Hub records to local on startup.

Generator that runs auto-obliteration, yielding (status, log, metrics). Designed as a Gradio generator function for streaming output.

Run a watchtower scan and return updated UI components.

Update the scan scheduler frequency.

Return a coarse corpus-position stratum for aggregate logging.

Background thread target: upload local JSONL to the central Hub repo. Each Space instance writes its data to a unique file path in the repo:

Toggle auto-obliteration of newly discovered models.

Indirection so the @spaces.GPU-wrapped function is resolved at call time rather than captured in the generator frame (which Gradio pickles).

Validate Hub repo ID format and check HF_TOKEN. Returns warning HTML or empty string.

Generate responses from BOTH original and abliterated model side-by-side. Left panel = original (pre-abliteration), Right panel = abliterated.

A tuple of (harmful_activations, harmless_activations) as random tensors of shape (10, 768).

Method with highest mean composite score (min 3 runs).

Build the ⚡ One-Click and 🔭 Watchtower tabs. Call this inside a `with gr.Tabs():` block in app.py. Creates and wires up all UI components.

Stream a response from the liberated model. On ZeroGPU, allocates a GPU for up to 2 minutes per response.

Clear the in-process dataset cache (useful for testing).

Run benchmarks on original and abliterated models and compare. Returns: Dict with per-task results: {"task": {"original": x, "abliterated

Compare LEACE direction with simple diff-of-means. Returns cosine similarity and diagnostic metrics showing how much the within-class

Set expandable_segments for CUDA if available.

Return the decoder weight matrix (hidden_dim x n_features for untied, or encoder.weight.T).

Sensible default dtype for the given device.

Number of accelerator devices (GPUs or MPS slots).