TrevorS

TrevorS / gemma-4-abliteration

Public

Gemma 4 abliteration research: biprojection + EGA for E2B, E4B, 26B MoE, 31B

19
1
100% credibility
Found Apr 08, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This repository offers scripts and techniques to modify Google's Gemma 4 models, reducing their refusal rates on sensitive prompts while preserving overall performance.

How It Works

1
🔍 Discover Uncensored AI

You hear about a way to make Google's Gemma AI models more open and willing to chat about anything without refusing.

2
📖 Read the Guide

Visit the project page to see simple examples, results showing fewer refusals, and links to ready-made models.

3
⚙️ Prepare Your Setup

Follow easy steps to get the needed tools on your computer so you can process the models yourself.

4
✨ Uncensor a Model

Pick a model size like small or powerful, then run the quick process to remove the built-in restrictions.

5
📊 Test the Results

Try harmful questions and see how it now answers freely with almost no refusals, just like shown in the charts.

6
💾 Save and Share

Save your new open model or grab pre-made ones from the shared links to use anywhere.

🎉 Chat Freely

Now enjoy an AI assistant that helps with any topic without holding back, feeling more natural and useful.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is gemma-4-abliteration?

This Python repo applies abliteration techniques like biprojection and EGA to Google's Gemma 4 models on Hugging Face—E2B, E4B dense models, plus 26B MoE and 31B—stripping out refusal behaviors from safety training. Developers run simple CLI scripts with PyTorch and Transformers to process any Gemma checkpoint, evaluate on 686 harmful prompts across datasets, and export uncensored bf16 weights or GGUF quants for llama.cpp. Result: models with sub-1% refusals and minimal KL divergence shift, ready for local inference.

Why is it gaining traction?

Unlike generic Gemma GitHub code or cookbook tweaks, it handles the full Gemma 4 family including tricky MoE layers via EGA, delivering audited low-refusal models directly to HF repos like TrevorJS/gemma-4-E2B-it-uncensored. Pre-built GGUF exports bypass heavy compute needs, and cross-eval dashboards prove generalization beyond toy benchmarks. Devs grab it for plug-and-play uncensored Gemma without reinventing Gemma tokenizer or Transformers hacks.

Who should use this?

ML engineers deploying local LLMs via llama.cpp who want Gemma's speed without refusals on creative or red-team tasks. Fine-tuners starting from DeepMind Gemma GitHub bases like 26B or 31B for custom uncensored agents. Teams bypassing Gemma 3 abliterated leftovers for fresh E2B/E4B variants in PyTorch pipelines.

Verdict

Solid starter for Gemma obliteration experiments—19 stars and 1.0% credibility reflect early maturity, but thorough README and HF exports make models usable today without running scripts. Fork if tweaking biprojection; otherwise, download the GGUF for immediate wins.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.