Zero-training, in-place weight surgery and directional abliteration engine for 4-bit quantized LLMs using Winsorized activation profiling and Gram-Schmidt orthogonal projection
research transformers pytorch quantization ai-safety gram-schmidt mechanistic-interpretability llm-security bitsandbytes representation-engineering llama3 qwen2-5 activation-steering abliteration weight-surgery
-
Updated
Nov 18, 2026 - Jupyter Notebook