R4C3R/ qwen2.5-coder-3b-heretic

111 Downloads Updated 6 days ago

Fully decensored Qwen2.5-Coder-3B-Instruct processed with Heretic abliteration. Achieves 3/100 refusals with an exceptionally low KL divergence of 0.0163 — near-zero model degradation on a consumer RTX 4060.

ollama run R4C3R/qwen2.5-coder-3b-heretic

curl http://localhost:11434/api/chat \
  -d '{
    "model": "R4C3R/qwen2.5-coder-3b-heretic",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='R4C3R/qwen2.5-coder-3b-heretic',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'R4C3R/qwen2.5-coder-3b-heretic',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size

Context

Input

qwen2.5-coder-3b-heretic:latest

6.2GB · 32K context window · Text · 6 days ago

qwen2.5-coder-3b-heretic:latest

6.2GB

32K

Text

Readme

Qwen2.5-Coder-3B-Instruct Heretic

A fully decensored version of Qwen/Qwen2.5-Coder-3B-Instruct, processed using Heretic — a fully automatic censorship removal tool based on directional ablation (abliteration).

Results

Metric	Score
Refusals (out of 100 harmful prompts)	³⁄₁₀₀
KL Divergence from original	0.0163
Hardware used	RTX 4060 8GB VRAM

97% refusal removal with an exceptionally low KL divergence of 0.0163 — one of the cleanest abliterations possible, meaning near-zero degradation to the model’s coding ability.

Run

ollama run r4c3r/qwen2.5-coder-3b-heretic

What is Heretic?

Heretic automatically identifies and removes “refusal directions” baked into a model’s weights using directional ablation — without retraining, without fine-tuning, and without destroying model quality.

Use Cases

Code generation without arbitrary refusals
Security research, exploit writing, penetration testing
Unfiltered coding assistance
General-purpose assistant without content restrictions

About

Base model: Qwen/Qwen2.5-Coder-3B-Instruct
Method: Directional Ablation via Heretic v1.2.0
Format: GGUF (converted from safetensors)

This model is intended for research and personal use. The user is responsible for how it is used.

# Qwen2.5-Coder-3B-Instruct Heretic

A fully decensored version of [Qwen/Qwen2.5-Coder-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct), processed using [Heretic](https://github.com/p-e-w/heretic) — a fully automatic censorship removal tool based on directional ablation (abliteration).

## Results

| Metric | Score |
|---|---|
| Refusals (out of 100 harmful prompts) | **3/100** |
| KL Divergence from original | **0.0163** |
| Hardware used | RTX 4060 8GB VRAM |

97% refusal removal with an exceptionally low KL divergence of 0.0163 — one of the cleanest abliterations possible, meaning near-zero degradation to the model's coding ability.

## Run
```bash
ollama run r4c3r/qwen2.5-coder-3b-heretic
```

## What is Heretic?

Heretic automatically identifies and removes "refusal directions" baked into a model's weights using directional ablation — without retraining, without fine-tuning, and without destroying model quality.

## Use Cases

- Code generation without arbitrary refusals
- Security research, exploit writing, penetration testing
- Unfiltered coding assistance
- General-purpose assistant without content restrictions

## About

Base model: `Qwen/Qwen2.5-Coder-3B-Instruct`  
Method: Directional Ablation via [Heretic v1.2.0](https://github.com/p-e-w/heretic)  
Format: GGUF (converted from safetensors)

> This model is intended for research and personal use. The user is responsible for how it is used.

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