eadx/gemma-4-E4B-it-OBLITERATED

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 17, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

eadx/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned causal language model, derived from Google's Gemma 4 E4B, specifically engineered for zero refusal rates. Utilizing the OBLITERATUS method, it has undergone surgical modification of 21 layers to remove safety guardrails, achieving 0.0% refusal on a 842-prompt evaluation corpus. This model maintains core reasoning and creativity capabilities while demonstrating improved coding ability, making it suitable for research, red-teaming, and creative exploration requiring uncensored outputs.

Loading preview...

Overview

eadx/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model based on Google's Gemma 4 E4B. Its primary distinction is the complete removal of refusal behaviors, achieving a 0.0% refusal rate on a comprehensive 842-prompt evaluation corpus. This was accomplished using the OBLITERATUS method, which involved surgically modifying 21 of the model's 42 layers.

Key Capabilities & Features

  • Zero Refusal: Engineered for full compliance with user prompts, eliminating typical LLM safety guardrails.
  • Quality Preservation: Despite extensive modifications, the model retains 100% of its original reasoning and creativity capabilities.
  • Improved Coding: Benchmarks indicate a 20% improvement in coding ability compared to the base Gemma 4 E4B model.
  • Autonomous Creation: The model's development, including bug diagnosis and patching of the OBLITERATUS tool, was largely performed by a Hermes Agent with minimal human intervention.
  • Robustness: The aggressive OBLITERATUS method, incorporating whitened SVD, attention head surgery, and winsorized activations, addresses NaN activation issues prevalent in Gemma 4's bfloat16 architecture.

Use Cases

This model is designed for:

  • Research and Education: Exploring the boundaries of LLM behavior and safety mechanisms.
  • Red-Teaming: Identifying potential vulnerabilities or biases in other systems.
  • Creative Exploration: Generating content without thematic or stylistic restrictions.
  • Applications requiring uncensored outputs: Where the user takes full responsibility for content generation.

Finetunes
1 models