Modelpublic

signsur4739379373/Qwen3.6-27B-Heretic-LoRA

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes114downloads

Model Card

Qwen3.6-27B-Heretic-LoRA

This is a standalone, lightweight LoRA adapter for Qwen3.6-27B that applies Heretic-style abliteration (uncensoring).

Please be mindful: The abliteration capabilities of the LoRA is sub-par to the actual Heretic checkpoint. If you're looking for proper abliteration, please use full weights, This LoRA exists to serve a technical need.

Instead of downloading an entirely new checkpoint, you can load this adapter directly on top of the stock Qwen3.6-27B base model (or any of its quantized variants) in llama.cpp to remove most refusal vectors.

Technical Details

—Method: SVD low-rank extraction of the orthogonal projection delta between the base model and the Heretic-abliterated weights (Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16).
—Target Modules: o_proj, down_proj, gate_proj, up_proj
—Rank: 16, 32 (recommended), 64, 128, 256
—Format: GGUF (qwen3.6-27b-heretic-lora-rank32.gguf)

Usage (llama.cpp)

Load your preferred quant of the base model, and attach the adapter using the --lora flag:

bash

llama-server -m qwen3.6-27b-Q4_K_M.gguf --lora qwen3.6-27b-heretic-lora-rank32.gguf