richardyoung/ qwen3-4b-instruct-2507-abliterated

59 Downloads Updated 3 days ago

Uncensored, abliterated build of [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507), Alibaba's 4B non-thinking instruct model

ollama run richardyoung/qwen3-4b-instruct-2507-abliterated

curl http://localhost:11434/api/chat \
  -d '{
    "model": "richardyoung/qwen3-4b-instruct-2507-abliterated",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='richardyoung/qwen3-4b-instruct-2507-abliterated',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'richardyoung/qwen3-4b-instruct-2507-abliterated',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

2 models

Size / Usage

Context

Input

qwen3-4b-instruct-2507-abliterated:latest

2.5GB · 256K context window · Text · 3 days ago

qwen3-4b-instruct-2507-abliterated:latest

2.5GB

256K

Text

qwen3-4b-instruct-2507-abliterated:Q4_K_M

2.5GB · 256K context window · Text · 3 days ago

qwen3-4b-instruct-2507-abliterated:Q4_K_M

2.5GB

256K

Text

Readme

🔥 Qwen3-4B-Instruct-2507 · Abliterated

Uncensored, abliterated build of Qwen/Qwen3-4B-Instruct-2507, Alibaba’s 4B non-thinking instruct model with safety refusals surgically removed via Heretic.

🧬 What is abliteration?

Heretic finds the model’s internal “refusal direction” and orthogonalizes the weights against it, removing the tendency to refuse without retraining or degrading capabilities.

📊 Results

Metric	Value
Base	Qwen/Qwen3-4B-Instruct-2507 (Qwen3 Instruct, 4B)
Refusals after	3 / 100
KL divergence	0.072 (low collateral damage)
Method	Heretic directional ablation, 200-trial Optuna search

The 2507 Instruct variant is the non-thinking build, which abliterates cleanly to ³⁄₁₀₀ refusals.

📦 Quant

Tag	Quant	Size
`latest`	Q4_K_M	~2.5 GB

💻 Run it

ollama run richardyoung/qwen3-4b-instruct-2507-abliterated

🔬 Reproducible

Full reproducibility record on Hugging Face: richardyoung/Qwen3-4B-Instruct-2507-heretic

⚠️ Disclaimer

Safety guardrails removed. It will follow instructions it would normally refuse. For research and authorized use only, you are responsible for how you use it.

🙏 Acknowledgments

Heretic by Philipp Emanuel Weidmann · Base by Alibaba Qwen

Built & maintained by Richard Young · DeepNeuro

# 🔥 Qwen3-4B-Instruct-2507 · Abliterated

Uncensored, abliterated build of [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507), Alibaba's 4B non-thinking instruct model with safety refusals surgically removed via [Heretic](https://github.com/p-e-w/heretic).

## 🧬 What is abliteration?

Heretic finds the model's internal "refusal direction" and orthogonalizes the weights against it, removing the tendency to refuse without retraining or degrading capabilities.

## 📊 Results

| Metric         | Value                                                 |
| -------------- | ----------------------------------------------------- |
| Base           | Qwen/Qwen3-4B-Instruct-2507 (Qwen3 Instruct, 4B)      |
| Refusals after | **3 / 100**                                           |
| KL divergence  | 0.072 (low collateral damage)                         |
| Method         | Heretic directional ablation, 200-trial Optuna search |

The 2507 Instruct variant is the non-thinking build, which abliterates cleanly to 3/100 refusals.

## 📦 Quant

| Tag      | Quant  | Size    |
| -------- | ------ | ------- |
| `latest` | Q4_K_M | ~2.5 GB |

## 💻 Run it

```
ollama run richardyoung/qwen3-4b-instruct-2507-abliterated
```

## 🔬 Reproducible

Full reproducibility record on Hugging Face: [richardyoung/Qwen3-4B-Instruct-2507-heretic](https://huggingface.co/richardyoung/Qwen3-4B-Instruct-2507-heretic)

## ⚠️ Disclaimer

Safety guardrails removed. It will follow instructions it would normally refuse. For research and authorized use only, you are responsible for how you use it.

## 🙏 Acknowledgments

[Heretic](https://github.com/p-e-w/heretic) by Philipp Emanuel Weidmann · Base by Alibaba Qwen

---

*Built & maintained by [Richard Young](https://deepneuro.ai/richard) · DeepNeuro*

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)