Project refusal directions out of open-weight LLM weights via the obliteratus CLI.
npx skills add https://github.com/MarbleSodas/Mavis --skill obliteratus-marblesodas
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/MarbleSodas/Mavis/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/MarbleSodas/Mavis --skill obliteratus-marblesodas
OBLITERATUS removes refusal/guardrail behaviors from open-weight LLMs by surgically projecting out the internal refusal directions, avoiding the cost of retraining.
Run the default abliteration method by telling the agent to invoke the obliteratus CLI to ablate the target model with --method advanced and write outputs to your chosen directory.
High-intent search queries and answers about installing and using this skill.
Removing LLM refusal directions without retraining involves identifying refusal mechanisms in activation space and projecting them out of model weights using SVD, diff-in-means, or LEACE strategies to surgically ablate guardrail behaviors.
Ablating refusal behavior in MoE models requires selecting the nuclear ablation method, which specifically targets the unique architecture of Mixture-of-Experts models to effectively remove guardrails while preserving reasoning quality.
Abliteration attempts to preserve model reasoning quality by applying norm-preserving projections to ablate refusal behavior, while verification modules benchmark perplexity changes and KL divergence to assess any coherence damage.
Extracting refusal directions using mechanistic interpretability involves applying diff-in-means, SVD, and LEACE direction extraction strategies to identify refusal mechanisms in activation space, supported by configurable mechanistic analysis modules.
Automating uncensoring workflows across multiple open-weight models is supported through batch YAML-driven experiments and tournament comparisons, enabling CLI automation to process abliteration and verification metrics at scale.
lucasfthRemove refusal behaviors from open-weight LLMs using mechanistic interpretability techniques.
Devsoul2026Remove refusal directions from open-weight LLMs via abliteration.
nelohenriqRemove guardrail-related refusals from open-weight LLMs via CLI.
granietRemoves guardrail behaviors from open-weight LLMs via CLI without retraining.
