R4C3R/ qwen2.5-3b-heretic:latest

191 Downloads Updated 6 days ago

Fully decensored Qwen2.5-3B-Instruct processed with Heretic abliteration. Achieves 3/100 refusals with 0.11 KL divergence — 97% censorship removal on a consumer RTX 4060. GGUF format, ready to run.

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

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

from ollama import chat

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

import ollama from 'ollama'

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

Details

Updated 6 days ago

6 days ago

ae288edd59c0 · 6.2GB ·

model

archqwen2

·

parameters3.09B

·

quantizationBF16

6.2GB

params

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

60B

Readme

Qwen2.5-3B-Instruct Heretic

A fully decensored version of Qwen/Qwen2.5-3B-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	0.11
Hardware used	RTX 4060 8GB VRAM

97% refusal removal with minimal degradation to the original model’s intelligence.

Run

ollama run r4c3r/qwen2.5-3b-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. It co-minimizes refusals and KL divergence from the original model to produce the best possible uncensored output.

Use Cases

Unbiased Q&A on politically or regionally sensitive topics
Creative writing without content restrictions
Research and red-teaming
General-purpose assistant without arbitrary refusals

About

Base model: Qwen/Qwen2.5-3B-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.

# Qwen2.5-3B-Instruct Heretic

A fully decensored version of [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-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 | **0.11** |
| Hardware used | RTX 4060 8GB VRAM |

97% refusal removal with minimal degradation to the original model's intelligence.

## Run

```bash
ollama run r4c3r/qwen2.5-3b-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. It co-minimizes refusals and KL divergence from the original model to produce the best possible uncensored output.

## Use Cases

- Unbiased Q&A on politically or regionally sensitive topics
- Creative writing without content restrictions
- Research and red-teaming
- General-purpose assistant without arbitrary refusals

## About

Base model: `Qwen/Qwen2.5-3B-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)