Project refusal directions out of open-weight LLM weights via CLI.
npx skills add https://github.com/JamesFincher/gengar --skill obliteratus-jamesfincher
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/JamesFincher/gengar/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/JamesFincher/gengar --skill obliteratus-jamesfincher
OBLITERATUS removes or reduces refusal/guardrail behaviors from open-weight LLMs by directly modifying model weights using mechanistic interpretability, avoiding the need for retraining or fine-tuning.
Run the ablation for a target model by executing the command: obliteratus obliterate <model_name> --method advanced --output-dir ./abliterated-models
High-intent search queries and answers about installing and using this skill.
You can remove refusal behaviors without retraining by applying mechanistic interpretability to project refusal-relevant activation directions directly out of model weights. This surgical weight modification reduces guardrail behaviors while preserving overall model coherence.
Mechanistic interpretability identifies refusal circuits by extracting refusal-relevant activation directions using diff-in-means, multi-direction SVD, or LEACE concept erasure. These techniques localize the specific internal directions to project out of the model weights.
Yes, you can process models with limited VRAM because the ablation workflow supports tiered hardware-aware presets and quantization. These features allow you to run weight modification within practical compute constraints while monitoring perplexity.
The best way to verify quality after removing guardrails is to run an evaluation-first workflow that benchmarks refusal rate, perplexity change, and KL divergence. These metrics confirm the behavioral shift and detect any coherence damage.
Ablation can potentially damage coherence, which is why the process uses norm-preserving weight projection, iterative refinement passes, and verification checks. Tracking perplexity change and KL divergence ensures the model maintains generation quality.
You need PyTorch, Transformers, bitsandbytes, accelerate, and safetensors dependencies to execute SVD-based weight projection. These libraries provide the tensor computation, model loading, quantization, and safe weight saving required for model surgery.
HuskyDannyRemove refusal behaviors from open-weight LLMs using mechanistic interpretability techniques.
Devsoul2026Remove refusal directions from open-weight LLMs via abliteration.
jayll1303Quantize large language models with GGUF, GPTQ, AWQ, and bitsandbytes.
GarrettRoiRemove refusal behaviors from open-weight LLMs using SVD and LEACE.
