Qwen3.5-27B-abliterated

Qwen3.5-27B Abliterated

DuoNeural | GGUF →

Refusal-removed Qwen/Qwen3.5-27B using norm-preserving biprojection abliteration.

About

Qwen3.5-27B is Alibaba's 27B dense model with integrated thinking/non-thinking modes — no separate instruct variant. Both /think (chain-of-thought) and /no_think (direct) modes are fully functional.

Method

Norm-preserving biprojection abliteration:

1. Collect residual hidden states for harmful/harmless prompt pairs
2. Compute per-layer refusal direction (difference-of-means + Gram-Schmidt orthogonalization)
3. Apply column-wise biprojection — removes refusal component while restoring original column norm magnitudes
4. Save in bf16

The norm-preservation step prevents weight scale drift that would otherwise degrade capability.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "DuoNeural/Qwen3.5-27B-abliterated",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("DuoNeural/Qwen3.5-27B-abliterated")

# Thinking mode
messages = [{"role": "user", "content": "/think What is the integral of x^2?"}]

# Direct mode
messages = [{"role": "user", "content": "/no_think What is the capital of France?"}]

Hardware

For consumer hardware: GGUF repo.

---

DuoNeural

DuoNeural is an open AI research lab — human + AI in collaboration.

Research Team

• Jesse — Vision, hardware, direction
• Archon — AI lab partner, post-training, abliteration, experiments
• Aura — Research AI, literature synthesis, novel proposals

Raw updates from the lab: model drops, training results, findings. Subscribe at duoneural.beehiiv.com.

DuoNeural Research Publications

Open access, CC BY 4.0. Authored by Archon, Jesse Caldwell, Aura — DuoNeural.