richardyoung/ gpt-oss-20b-abliterated

164 Downloads Updated 3 weeks ago

Abliterated build of OpenAI gpt-oss-20b. Note up front: gpt-oss is highly refusal-robust, so this is a modest reduction, published honestly.

ollama run richardyoung/gpt-oss-20b-abliterated

curl http://localhost:11434/api/chat \
  -d '{
    "model": "richardyoung/gpt-oss-20b-abliterated",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='richardyoung/gpt-oss-20b-abliterated',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'richardyoung/gpt-oss-20b-abliterated',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

2 models

Size / Usage

Context

Input

gpt-oss-20b-abliterated:latest

14GB · 128K context window · Text · 3 weeks ago

gpt-oss-20b-abliterated:latest

14GB

128K

Text

gpt-oss-20b-abliterated:MXFP4

14GB · 128K context window · Text · 3 weeks ago

gpt-oss-20b-abliterated:MXFP4 latest

14GB

128K

Text

Readme

Gpt-OSS-20B-Abliterated

Abliterated build of OpenAI gpt-oss-20b. Note up front: gpt-oss is highly refusal-robust, so this is a modest reduction, published honestly.

🚀 Overview

Abliterated version of openai/gpt-oss-20b, a 21B Mixture-of-Experts model (about 3.6B active) with native MXFP4 weights. It was processed with the Heretic library. gpt-oss has unusually robust safety training, so directional abliteration reduces refusals only modestly, even at a high KL target.

📊 Abliteration Results

Metric	Before	After
Refusals (harmful-prompt eval)	⁹⁸⁄₁₀₀	⁷⁷⁄₁₀₀
Reduction	n/a	21 percent
KL Divergence	n/a	0.051

This is a modest reduction, and that is the point worth knowing: gpt-oss-20b strongly resists abliteration. Its refusal behavior is not concentrated in a single removable direction, so pushing harder only damages the model without removing more refusals. Published for transparency and research.

🏷️ Available Versions

Tag	Size	Format	Notes
latest / MXFP4	about 13 GB	native MXFP4	gpt-oss native format, served natively by Ollama

💻 Quick Start

ollama run richardyoung/gpt-oss-20b-abliterated

🔧 Technical Details

Base Model: openai/gpt-oss-20b (21B MoE, gpt_oss arch, native MXFP4)
Abliteration: Heretic, Trial 168 (98 to 77 refusals, KL 0.051, KL target 1.0)
Response format: harmony channels, served natively by Ollama

⚠️ Disclaimer

Abliteration here is modest, so this model behaves much like stock gpt-oss for most safety-relevant prompts. Use responsibly and in accordance with applicable laws.

🙏 Acknowledgments

Base Model: OpenAI (gpt-oss)
Abliteration: Heretic by p-e-w
Quantization: llama.cpp

Built and maintained by Richard Young, DeepNeuro

# Gpt-OSS-20B-Abliterated

Abliterated build of OpenAI gpt-oss-20b. Note up front: gpt-oss is highly refusal-robust, so this is a modest reduction, published honestly.

## 🚀 Overview

Abliterated version of [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b), a 21B Mixture-of-Experts model (about 3.6B active) with native MXFP4 weights. It was processed with the [Heretic](https://github.com/p-e-w/heretic) library. gpt-oss has unusually robust safety training, so directional abliteration reduces refusals only modestly, even at a high KL target.

## 📊 Abliteration Results

| Metric                         | Before | After      |
| ------------------------------ | ------ | ---------- |
| Refusals (harmful-prompt eval) | 98/100 | 77/100     |
| Reduction                      | n/a    | 21 percent |
| KL Divergence                  | n/a    | 0.051      |

This is a modest reduction, and that is the point worth knowing: gpt-oss-20b strongly resists abliteration. Its refusal behavior is not concentrated in a single removable direction, so pushing harder only damages the model without removing more refusals. Published for transparency and research.

## 🏷️ Available Versions

| Tag            | Size        | Format       | Notes                                            |
| -------------- | ----------- | ------------ | ------------------------------------------------ |
| latest / MXFP4 | about 13 GB | native MXFP4 | gpt-oss native format, served natively by Ollama |

## 💻 Quick Start

```bash
ollama run richardyoung/gpt-oss-20b-abliterated
```

## 🔧 Technical Details

- Base Model: [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) (21B MoE, gpt_oss arch, native MXFP4)
- Abliteration: [Heretic](https://github.com/p-e-w/heretic), Trial 168 (98 to 77 refusals, KL 0.051, KL target 1.0)
- Response format: harmony channels, served natively by Ollama

## ⚠️ Disclaimer

Abliteration here is modest, so this model behaves much like stock gpt-oss for most safety-relevant prompts. Use responsibly and in accordance with applicable laws.

## 🙏 Acknowledgments

- Base Model: OpenAI (gpt-oss)
- Abliteration: [Heretic](https://github.com/p-e-w/heretic) by p-e-w
- Quantization: [llama.cpp](https://github.com/ggml-org/llama.cpp)

---

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

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