skills.rest
xu1713
xu1713COMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Removes refusal behaviors from open-weight LLMs via weight projection and SVD.

Updated Jun 5, 2026
Search Tags:#llm#svd#huggingface#mechanistic interpretability#model surgery#weight projection#abliteration
One-click install
npx skills add https://github.com/xu1713/openhorse --skill obliteratus-xu1713
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/xu1713/openhorse/tree/main/openhorse/openhorse/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/xu1713/openhorse --skill obliteratus-xu1713

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires obliteratus, torch, transformers, bitsandbytes, accelerate, safetensors, and includes references (resource) components.

What problem does it solve? This Skill modifies open-weight language models to remove refusal and guardrail behaviors without retraining or fine-tuning, using mechanistic interpretability techniques to identify and excise refusal directions from model weights while preserving reasoning capabilities. ## Core Features & Use Cases - Nine CLI abliteration methods: Choose from basic, advanced, aggressive, surgical, nuclear, and other methods matched to model architecture (dense, MoE, reasoning models) and quality/speed tradeoffs. - 28 analysis modules: Fingerprint alignment training methods (DPO/RLHF/CAI/SFT), map refusal cone geometry, trace causal components, and detect self-repair tendencies before modifying weights. - Hardware-aware workflow: Check GPU VRAM, browse 116 model presets across 5 compute tiers, and get telemetry-driven method recommendations. - Use Case: A researcher wants to study how refusal is encoded in Llama 3.1 8B. They run analysis modules to locate the refusal direction, apply the advanced method with 4-bit quantization, then verify the refusal rate dropped below 5% while perplexity increased less than 10%. ## Quick Start Ask the agent to check GPU availability, install OBLITERATUS, and run the advanced abliteration method on a chosen HuggingFace model with 4-bit quantization.

Frequently Asked Questions about obliteratus

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I remove refusals from an LLM without fine-tuning?▼

Use abliteration to extract a refusal direction from model activations and project it out of the weight matrices. Run the OBLITERATUS CLI with the advanced method, which applies multi-direction SVD with norm-preserving projection in about 10-20 minutes for an 8B model.

Which abliteration method should I use for my model?▼

Use advanced as the default for most dense models like Llama and Mistral. Choose nuclear for MoE models like Mixtral or DeepSeek, surgical for reasoning models like R1 distills to preserve chain-of-thought, and aggressive only if advanced leaves more than 10% refusals.

How much VRAM does abliteration require?▼

With 4-bit quantization, 8-16 GB VRAM handles models up to 9B parameters, 24 GB handles up to 32B, and 48 GB or more handles 72B models. CPU-only machines are limited to models under 1B parameters.

Why does my abliterated model produce repetitive or incoherent output?▼

Coherence damage occurs when too many directions are removed or regularization is too low, shown by perplexity increasing over 15%. Reduce n-directions, increase regularization to 0.3, lower refinement passes to 1, or retry with the gentler basic method.

Can I import OBLITERATUS as a Python library in my project?▼

No, OBLITERATUS is AGPL-3.0 licensed, so importing it as a library would impose AGPL terms on your project. Invoke it only via the CLI or subprocess to keep MIT or Apache licensed projects clean.

Does abliteration work on small models under 1B parameters?▼

Models under 1B respond poorly because their refusal behaviors are shallow and fragmented, making clean direction extraction difficult. Expect 20-40% residual refusal; models 3B and larger have cleaner refusal directions and often reach 0% refusal.

Related Skills

Explore More Software Engineering Skills→
danmartinez78danmartinez78

obliteratus

Remove refusal directions from open-weight LLMs using mechanistic interpretability techniques.

Community
Advanced
GitnappGitnapp

gguf-quantization

Convert Hugging Face models into GGUF files with Q4_K_M quantization.

Community
Advanced
qcmuuqcmuu

evaluating-llms-harness

Run 60+ academic benchmarks on LLMs with lm-evaluation-harness.

Community
Intermediate
cuba6112cuba6112

unsloth-fft

Perform full fine-tuning on large language models with Unsloth libraries.

Community
Advanced