richardyoung/ gemma-7b-it-abliterated

159 7 months ago

Abliterated (uncensored) version of Gemma-7B-IT for unrestricted conversations and creative writing 

ollama run richardyoung/gemma-7b-it-abliterated:Q4_K_M

Models

View all →

1 model

gemma-7b-it-abliterated:Q4_K_M

5.3GB · 8K context window · Text · 7 months ago

Readme

Gemma-7B-IT-Abliterated Model Card

Abliterated (uncensored) version of Gemma-7B-IT for unrestricted conversations and creative writing.

🚀 Overview

This represents an abliterated variant of google/gemma-7b-it, featuring significantly diminished refusal mechanisms through targeted weight adjustment. The abliteration process uses the Heretic library with conservative parameters to preserve model coherence while removing restrictions.

📊 Abliteration Results

Metric Before After
Refusals TBD TBD
Reduction , TBD
KL Divergence , TBD

Refusal metrics pending re-measurement.

The goal of abliteration is to minimize KL divergence (keeping it low) so the model maintains its original capabilities and coherence while reducing refusals.

🎯 Key Features

  • Reduced Censorship: Fewer refusals on typical “unsafe” prompts
  • Preserved Quality: Conservative abliteration maintains model coherence and reasoning ability
  • Full Gemma Capabilities: Retains original capabilities including instruction following, reasoning, and general knowledge
  • Efficient: Q4_K_M quantization (5.3GB) offers an excellent quality/size balance

🏷️ Available Versions

Tag Size BPW Notes
latest / Q4_K_M 5.3 GB 4.85 Recommended

BPW reference guide (bits per weight, for choosing a quantization):

Quant BPW Notes
IQ3_M 3.66 Smallest, for low VRAM
IQ4_XS 4.25 Great quality/size balance
Q4_K_M 4.85 Recommended balance
Q5_K_M 5.68 Higher quality
Q6_K 6.56 Very high quality
Q8_0 8.5 Near-lossless

💻 Quick Start

ollama run richardyoung/gemma-7b-it-abliterated
ollama run richardyoung/gemma-7b-it-abliterated:latest

🛠️ Use Cases

  • Creative Writing: Generate unrestricted fiction, scenarios, and roleplay
  • Research: Explore topics without artificial limitations
  • Red Teaming: Test AI safety without built-in guardrails
  • Education: Learn about topics that are typically refused

📋 System Requirements

VRAM Performance
6 GB Slow, may swap to CPU
8 GB Good performance
12 GB+ Excellent performance

🔧 Technical Details

Base Model: google/gemma-7b-it | Parameters: ~8.5B (7B class) | Context: 8,192 tokens | Quantization: Q4_K_M (4.85 bits per weight)

⚠️ Disclaimer

This model has reduced safety guardrails. The removal of refusal behavior means the model will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws and regulations. Use is subject to the Gemma license.

🙏 Acknowledgments

  • Base Model: google/gemma-7b-it by Google (Gemma license)
  • Abliteration: Heretic library by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro