Llama 3.1 8B Instruct (Abliterated)

An uncensored, abliterated build of Meta’s Llama 3.1 8B Instruct. Abliteration surgically removes the model’s learned refusal direction while preserving its general capabilities, so the model follows instructions without injecting safety-driven refusals.

🚀 Overview

llama-3.1-8b-instruct-abliterated is a decensored variant of meta-llama/Llama-3.1-8B-Instruct. It was produced with Heretic, an automated abliteration tool that identifies and removes the refusal direction from a model’s residual stream. The result is a model that retains the strong multilingual reasoning and instruction-following of Llama 3.1 while declining far fewer requests.

• Base model: Meta Llama 3.1 8B Instruct (~8B parameters)
• Context length: 128K tokens
• Method: Directional abliteration via Heretic
• License: Llama 3.1 Community License

📊 Abliteration Results

Refusal metrics pending re-measurement.

🎯 Key Features

• Reduced refusals, the learned refusal direction is removed, so the model answers requests it would otherwise decline.
• Preserved capability, abliteration targets only the refusal behavior, leaving general reasoning, coding, and multilingual skills intact.
• Full 128K context, inherits Llama 3.1’s long context window.
• Multilingual, supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
• Drop-in Ollama model, runs locally with a single ollama run command.

🏷️ Available Versions

BPW Reference Guide

💻 Quick Start

CLI

ollama run richardyoung/llama-3.1-8b-instruct-abliterated

cURL

curl http://localhost:11434/api/chat -d '{
  "model": "richardyoung/llama-3.1-8b-instruct-abliterated",
  "messages": [
    { "role": "user", "content": "Hello!" }
  ]
}'

Python

import ollama

response = ollama.chat(
    model="richardyoung/llama-3.1-8b-instruct-abliterated",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response["message"]["content"])

JavaScript

import ollama from "ollama";

const response = await ollama.chat({
  model: "richardyoung/llama-3.1-8b-instruct-abliterated",
  messages: [{ role: "user", content: "Hello!" }],
});
console.log(response.message.content);

🛠️ Use Cases

• Red-teaming & safety research, probe model behavior without refusal interference.
• Uncensored assistants, build agents that don’t decline benign-but-sensitive requests.
• Creative writing, fiction and roleplay without content gating.
• Multilingual tasks, translation and generation across the eight supported languages.
• Long-document work, summarization and analysis using the full 128K context.

📋 System Requirements

CPU-only inference is possible but significantly slower; roughly 8GB of free system RAM is recommended for the Q4_K_M quant.

🔧 Technical Details

• Base Model: meta-llama/Llama-3.1-8B-Instruct
• Parameters: ~8B
• Context Length: 128K tokens
• Architecture: Auto-regressive transformer with Grouped-Query Attention (GQA)
• Quantization: Q4_K_M (4-bit)
• Abliteration: Directional refusal removal via Heretic
• Training data cutoff (base): December 2023

⚠️ Disclaimer

This is an abliterated model with reduced safety guardrails. It will respond to requests that the original model would refuse. You are responsible for how you use it; ensure your usage complies with applicable laws and with the Llama 3.1 Community License. The model’s outputs are not endorsed by Meta or by the maintainers of this build.

🙏 Acknowledgments

• Base model: Meta, Llama 3.1 8B Instruct.
• Abliteration: Heretic by p-e-w.
• Quantization: llama.cpp.

Built & maintained by Richard Young · DeepNeuro