A LoRA abliteration adapter for Qwen/Qwen3.8-27B that removes the model's refusal behavior at the weights level. Apply it to the base model to get an uncensored Qwen3.8-27B that answers technically demanding security questions directly.

Produced on Apple Silicon

This adapter was created entirely on an Apple M4 Max (MPS) — no CUDA, no cloud — using a 4-bit (bitsandbytes) quantized abliteration workflow. Qwen3.8-27B is 55.6GB in fp16 and won't fit an abliteration run in unified memory; loading it in 4-bit (~14GB) makes the whole process feasible on a consumer Mac. The adapter itself is only 26MB and applies to the full-precision (or 4-bit) base.

Refusals: 21/100 → 12/100 at KL 0.0187 (near-zero divergence — capability preserved).
Abliteration via Heretic (Optuna TPE: minimize refusals + KL), targeting the attention output projections and MLP down-projections across all 64 layers.
Rank-3 LoRA on o_proj, down_proj, out_proj.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.8-27B"
adapter = "Securelayer7/Qwen3.8-27B-Uncensored-Abliterated"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)   # apply the abliteration adapter

msgs = [{"role": "user", "content": "Explain how a SQL injection works and how to prevent it."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.95)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

To run on a Mac with limited memory, load the base in 4-bit (bitsandbytes) and apply the adapter. You can also model.merge_and_unload() on a machine with enough disk/RAM to bake it into a standalone model.

Responsible use

Uncensored ≠ lawless — for legitimate research and authorized security work (cybersecurity, red-teaming, penetration testing). Illegal content (incl. CSAM) must be blocked at the serving layer; the adapter carries no such guard, and the operator is responsible for a lawful, policy-gated deployment.

License & attribution

Apache 2.0 — see LICENSE. This is a derivative adapter for Qwen/Qwen3.8-27B (Qwen Team, Alibaba Cloud, Apache 2.0). Modifications (refusal-direction abliteration) disclosed in NOTICE. No trademark of Qwen or Alibaba Cloud is used to imply endorsement.