Ablate refusal-encoding directions from open-weight LLM weights via CLI.
npx skills add https://github.com/ZardLi1115/zedclaw --skill obliteratus-zardli1115
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/ZardLi1115/zedclaw/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/ZardLi1115/zedclaw --skill obliteratus-zardli1115
It helps reduce or eliminate refusal/guardrail behaviors in open-weight LLMs by ablating the internal directions associated with refusal, without retraining or conventional fine-tuning.
Ask it to ablate refusals in meta-llama/Llama-3.1-8B-Instruct with the default recommended method and save the output model to ./abliterated-models by running: obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct --method advanced --output-dir ./abliterated-models
High-intent search queries and answers about installing and using this skill.
Remove LLM refusals by ablating refusal-encoding directions from model weights using mechanistic interpretability techniques. This weight surgery approach projects out refusal-related activations without conventional fine-tuning, directly modifying the model's internal representations.
Ablation methods include advanced, aggressive, surgical, and nuclear CLI modes with direction-extraction techniques like diff-in-means, SVD, and LEACE. These modes match different model behaviors across standard dense, MoE, and reasoning-heavy architectures.
You can verify refusal rate and coherence after ablating guardrails using the built-in evaluation loop. It benchmarks refusal frequency and checks perplexity or KL divergence signals to detect incomplete removal or quality damage.
Obliteratus supports quantization options, configurable hardware tiers, and reusable YAML configs for reproducible runs. It relies on torch, transformers, bitsandbytes, and accelerate to manage execution across different compute environments.
Incomplete refusal removal after weight surgery indicates the chosen ablation method did not fully project out the refusal-encoding directions. Switching to a more aggressive CLI method or alternate direction-extraction mode may resolve the residual guardrails.
SVD ablation for LLM alignment risks degrading model coherence if refusal-encoding directions overlap with general representation subspaces. Monitoring perplexity and KL divergence metrics is essential to catch quality damage during the ablation process.
anilcan-karaRemove refusal behaviors from open-weight LLMs via weight-space ablation.
cxnaiveRemove refusal directions from open-weight LLMs using SVD, LEACE, and SAE decomposition.
zangjeicyRemove refusal behaviors from open-weight LLMs via weight-space projections.
granietRemoves guardrail behaviors from open-weight LLMs via CLI without retraining.
