MyClawMyClaw
HomeUse CasesPricingSkillsResourcesCompany
HomeUse CasesPricingSkills
Sign In
HomeUse CasesPricingSkills
ResourcesCompareBlogCommunity
CompanyAboutSupport
Sign In
MyClawMyClaw

Your AI agent, already running.

Product

OpenClaw HostingHermes Agent HostingSlackUse CasesPricingAlibaba Cloud

Tools

Passport Photo GeneratorTattoo DesignVirtual Background GeneratorLinkedIn Headshot GeneratorParty Invitation GeneratorAlbum Cover GeneratorDating Photo Generator

Compre

vs dotsvs Claudevs ChatGPTvs OpenClawvs Hermes Agentvs Grok Botvs Hostinger

Resources

BlogUpdatesSkillsModelsSupportCommunityCapybara

© 2023 Cubo World Inc. All rights reserved.

TermsPrivacyCookieGDPRContact
← Back to Skills
AI & LLMs
🔬

Obliteratus

Remove refusal behaviors from open-weight LLMs without retraining, using mechanistic interpretability techniques like diff-in-means, SVD, and LEACE concept erasure.

by NousResearchv1.0.0
Connecting to VM...
Connecting to VM...
npx clawhub@latest install obliteratus
5Current Installs
📦
v1.0.0Version
View Source

Obliteratus Skill Overview

OBLITERATUS removes refusal behaviors from open-weight LLMs without retraining or fine-tuning, using mechanistic interpretability techniques including diff-in-means, SVD, LEACE concept erasure, and SAE decomposition. It surgically excises refusal directions from model weights while preserving reasoning capabilities, producing a standard HuggingFace model ready for local inference, vLLM serving, or Hub upload. With 9 CLI methods, 28 analysis modules, and telemetry-driven recommendations, it covers everything from quick prototyping on a laptop GPU to frontier-scale 200B+ multi-GPU runs.

How to Use It

Step 1: Run in your terminal or install this skill on MyClaw

npx clawhub@latest install obliteratus
or

Click the Install button at the top of this page for one-click setup

When to Use Obliteratus

Best Fit

  • You want to remove guardrails from an open-weight model (Llama, Qwen, Mistral, Gemma, DeepSeek, etc.) without fine-tuning or retraining.
  • You are researching how refusal behaviors are encoded in model weight space and want mechanistic interpretability tooling.
  • You need to abliterate a reasoning model (R1 distillations) while preserving chain-of-thought behavior.
  • You are working with MoE architectures (DeepSeek-MoE, Mixtral) that require expert-granular refusal removal.

When Not to Use

  • You are targeting models under ~1B parameters — refusal directions in very small models are fragmented and expect only partial results (20–40% remaining refusal rate).
  • You need reversible, inference-time control rather than permanent weight modification — consider steering vectors instead.
  • You only have access to closed-weight or API-only models; OBLITERATUS requires local access to model weights.

Key Features

9 Abliteration Methods

From basic (single diff-in-means direction, ~5–10 min for 8B) to optimized (Bayesian hyperparameter search via Optuna TPE) and surgical (SAE features + neuron masking + head surgery for reasoning models). The default advanced method uses multi-direction SVD with norm-preserving projection and is recommended for most models.

28 Mechanistic Interpretability Analysis Modules

Pre-abliteration analysis tools including alignment_imprint (fingerprints DPO/RLHF/CAI/SFT alignment type), logit_lens (identifies which layer decides to refuse), causal_tracing (finds causally necessary components), and anti_ouroboros (self-repair risk scoring).

Telemetry-Driven Recommendations

The obliteratus recommend <model> command draws on anonymized community run data to suggest the optimal method and parameters for a given architecture before you commit to a run. The --insights flag surfaces global cross-architecture rankings.

Tournament Evaluation

obliteratus tourney <model> runs all methods head-to-head and ranks them by refusal rate, perplexity change, and KL divergence, making it straightforward to identify the best approach for a specific model without manual iteration.

116 Model Presets Across 5 Compute Tiers

A curated model list spanning CPU-only tiny models through frontier 200B+ multi-GPU configurations, browsable with obliteratus models --tier <tier>. VRAM requirements and recommended quantization settings are included for each preset.

Multiple Execution Modes

Supports interactive wizard (obliteratus interactive), Gradio web UI (obliteratus ui), YAML-driven ablation studies (obliteratus run config.yaml), and direct CLI invocation, covering both ad-hoc runs and reproducible batch processing workflows.

Use Cases

Uncensoring Open-Weight Models

Produce abliterated variants of Llama, Qwen, Mistral, Gemma, and similar models for local inference or Hub upload. The advanced method handles most dense architectures reliably with near-zero remaining refusal rate on 3B+ models.

Alignment Research & Mechanistic Interpretability

Use the 28 analysis modules to map where and how refusal behaviors are encoded before modifying weights. causal_tracing, concept_geometry, and logit_lens give layer-level and component-level insight into a model's alignment imprint.

MoE Architecture Abliteration

The nuclear method applies expert-granular processing to handle DeepSeek-MoE, Mixtral, and other mixture-of-experts architectures where standard direction-based methods leave residual refusals.

Reasoning Model Abliteration

The surgical method combines SAE features, neuron masking, and head surgery with chain-of-thought awareness, preserving reasoning capability in R1 distillations and other CoT-heavy models while stripping refusal behavior.

Requirements

  • Linux or macOS
  • Python with pip; OBLITERATUS installed via pip install -e . from the GitHub repository
  • NVIDIA GPU with CUDA recommended; Apple Silicon (MLX backend) and CPU (tiny models only) also supported
  • 4–48+ GB VRAM depending on target model size; 4-bit quantization (--quantization 4bit) available to reduce requirements
  • git for cloning the repository
  • HuggingFace CLI (huggingface-cli) if uploading abliterated models to the Hub
  • License: OBLITERATUS is AGPL-3.0 — invoke via CLI or subprocess only; never import as a Python library in MIT/Apache-licensed projects
Connecting to VM...
npx clawhub@latest install obliteratus
5Current Installs
📦
v1.0.0Version
View Source

Reviews

0 reviews

Log in to write a review

No reviews yet. Be the first to share your experience!