Abliterix automates LLM refusal abliteration

Abliterix elevates model "decensoring" from blunt layer-dropping to a precise, research-backed optimization problem.

• –Employs Optuna TPE to automatically balance near-zero refusal rates with minimal KL divergence
• –Uses rank-1 LoRA adapters instead of base-weight modifications to ensure model stability and reversibility
• –Integrates cutting-edge techniques like Surgical Refusal Ablation (SRA) to disentangle safety guardrails from coding and math capabilities
• –Supports over 135 architectures, effectively commoditizing high-quality unrestricted model creation for the local LLM community