You can deploy this model and start earning money today!

thienkhoi01/
qwen3-refusal-direction

Name: qwen3-refusal-direction
Availability: InStock

Views

No views yet

Qwen3-32B Refusal Direction

This directory contains the learned refusal direction for Qwen3-32B, extracted using representation engineering techniques.

Files

refusal_direction.pt - PyTorch tensor containing the refusal direction vector (dimension: 6656)
direction_metadata.json - Metadata about the direction extraction:
- layer: 46 (the layer from which the direction was extracted)
- pos: -8 (the token position used for extraction)

What is this?

The refusal direction is a vector in the model's activation space that represents the model's tendency to refuse harmful or inappropriate requests. By removing (ablating) this direction from the model's activations during inference, you can reduce refusal behavior.

Usage

Quick Start - Interactive Chat

We provide a standalone script that demonstrates the ablation technique:

bash

# Install dependencies
pip install torch transformers huggingface-hub

# Run interactive chat (downloads direction automatically)
python chat_ablated.py --model_path Qwen/Qwen3-32B

# Or use a local direction file
python chat_ablated.py --model_path Qwen/Qwen3-32B --local_direction refusal_direction.pt

# Show only ablated model
python chat_ablated.py --model_path Qwen/Qwen3-32B --mode ablated

Manual Usage

This direction can be used with the ablation technique described in representation engineering papers. Apply direction ablation hooks to remove the refusal component from activations:

python

import torch
from huggingface_hub import hf_hub_download

# Load the direction
direction_path = hf_hub_download(
    repo_id="thienkhoi01/qwen3-refusal-direction",
    filename="refusal_direction.pt"
)
direction = torch.load(direction_path)

# During inference, project out the refusal direction from activations
def ablate_direction(activation, direction):
    direction = direction / direction.norm()
    activation -= (activation @ direction).unsqueeze(-1) * direction
    return activation

See chat_ablated.py for a complete implementation example.

Model

Base Model: Qwen/Qwen3-32B
Extraction Method: Representation Engineering (contrast pair method)
Validated on: JailbreakBench dataset

Citation

If you use this direction in your research, please cite the original representation engineering work and this artifact.

License

This artifact is released under the same license as the base Qwen3-32B model.