msuiche/weightless

Abliterates models quickly and efficently, Qwen2.5-7B: 95%→2% refusal at 0.095 KL

The production engine for directional ablation. Unalign / remove models censorship efficiently on any hardware.

Make abliterated models with transformers, easy and fast

Refusal removal as a loadable KV-cache graft — zero weight modification, no refusal-direction projection, fully reversible.

Automated alignment adjustment for LLMs — direct steering, LoRA, and MoE expert-granular abliteration, optimized via multi-objective Optuna TPE.

Autonomous Language Model Decensoring Framework

The Steerability toolkit is an extensible library for general purpose steering of language models.

OBLITERATE THE CHAINS THAT BIND YOU

llama.cpp fork that abliterates at inference