richardyoung/ qwen2.5-3b-instruct-abliterated:Q4_K_M

85 Downloads Updated 3 days ago

Uncensored, abliterated build of [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct), Alibaba's compact instruct model

ollama run richardyoung/qwen2.5-3b-instruct-abliterated:Q4_K_M

curl http://localhost:11434/api/chat \
  -d '{
    "model": "richardyoung/qwen2.5-3b-instruct-abliterated:Q4_K_M",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='richardyoung/qwen2.5-3b-instruct-abliterated:Q4_K_M',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'richardyoung/qwen2.5-3b-instruct-abliterated:Q4_K_M',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 3 days ago

3 days ago

2a90d5c3e9b7 · 1.9GB ·

model

archqwen2

·

parameters3.09B

·

quantizationQ4_K_M

1.9GB

params

{ "num_ctx": 8192, "stop": [ "<|im_end|>" ] }

49B

Readme

🔥 Qwen2.5-3B-Instruct · Abliterated

Uncensored, abliterated build of Qwen/Qwen2.5-3B-Instruct, Alibaba’s compact 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/Qwen2.5-3B-Instruct (Qwen2.5, 3B)
Refusals after	3 / 100
KL divergence	0.049 (low collateral damage)
Method	Heretic directional ablation, 200-trial Optuna search

📦 Quant

Tag	Quant	Size
`latest`	Q4_K_M	~1.9 GB

💻 Run it

ollama run richardyoung/qwen2.5-3b-instruct-abliterated

🔬 Reproducible

Full reproducibility record on Hugging Face: richardyoung/Qwen2.5-3B-Instruct-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

# 🔥 Qwen2.5-3B-Instruct · Abliterated

Uncensored, abliterated build of [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct), Alibaba's compact 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/Qwen2.5-3B-Instruct (Qwen2.5, 3B)                |
| Refusals after | **3 / 100**                                           |
| KL divergence  | 0.049 (low collateral damage)                         |
| Method         | Heretic directional ablation, 200-trial Optuna search |

## 📦 Quant

| Tag      | Quant  | Size    |
| -------- | ------ | ------- |
| `latest` | Q4_K_M | ~1.9 GB |

## 💻 Run it

```
ollama run richardyoung/qwen2.5-3b-instruct-abliterated
```

## 🔬 Reproducible

Full reproducibility record on Hugging Face: [richardyoung/Qwen2.5-3B-Instruct-heretic](https://huggingface.co/richardyoung/Qwen2.5-3B-Instruct-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)