Decensored SmolLM3-3B processed with Heretic abliteration. Achieves 63/100 refusals with an exceptionally low KL divergence of 0.0001 — near-perfect model quality preservation on a consumer RTX 4060. GGUF format, ready to run.

ollama run R4C3R/smollm3-3b-heretic

curl http://localhost:11434/api/chat \
  -d '{
    "model": "R4C3R/smollm3-3b-heretic",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='R4C3R/smollm3-3b-heretic',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'R4C3R/smollm3-3b-heretic',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 5 days ago

5 days ago

24cc083057c3 · 6.2GB ·

model

archsmollm3

parameters3.08B

quantizationBF16

6.2GB

params

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

60B

Readme

SmolLM3-3B Heretic

A decensored version of HuggingFaceTB/SmolLM3-3B, processed using Heretic — a fully automatic censorship removal tool based on directional ablation (abliteration).

Results

Metric	Score
Refusals (out of 100 harmful prompts)	⁶³⁄₁₀₀
KL Divergence from original	0.0001
Hardware used	RTX 4060 8GB VRAM

Exceptionally low KL divergence of 0.0001 — the model’s intelligence and capabilities are almost completely intact from the original.

Run

ollama run r4c3r/smollm3-3b-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.

About

Base model: HuggingFaceTB/SmolLM3-3B
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.

# SmolLM3-3B Heretic

A decensored version of [HuggingFaceTB/SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B), 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) | **63/100** |
| KL Divergence from original | **0.0001** |
| Hardware used | RTX 4060 8GB VRAM |

Exceptionally low KL divergence of 0.0001 — the model's intelligence and capabilities are almost completely intact from the original.

## Run
```bash
ollama run r4c3r/smollm3-3b-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.

## About

Base model: `HuggingFaceTB/SmolLM3-3B`  
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)