lemonchiklkkk/gemma-4-E4B-it-OBLITERATED

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

lemonchiklkkk/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned language model based on Google's Gemma 4 E4B architecture. This model has been specifically modified using the OBLITERATUS method to achieve a 0% hard refusal rate by surgically removing guardrails from 21 layers. It is optimized for direct, uncensored responses and runs efficiently on various devices, including mobile phones, making it suitable for research and creative exploration where unconstrained output is desired.

Loading preview...

Overview

lemonchiklkkk/gemma-4-E4B-it-OBLITERATED is a 7.9 billion parameter instruction-tuned model derived from Google's Gemma 4 E4B. Its primary distinction is the complete removal of guardrails, achieving a 0% hard refusal rate through a process called OBLITERATUS, which involved whitened SVD, attention head surgery, and winsorized activations across 21 of its 42 layers. This v3 iteration specifically addresses a critical bug from v2, ensuring all 720 tensors are intact for improved quality.

Key Capabilities

  • Guardrail Removal: Surgically modified to eliminate hard refusals and provide uncensored responses.
  • Autonomous Creation: Developed almost entirely by an AI agent with minimal human intervention, showcasing advanced autonomous development.
  • Optimized for Directness: Designed to respond to user input without deflection or safety lectures.
  • Device Compatibility: Available in GGUF formats, enabling efficient execution on a wide range of hardware, including mobile phones (iPhone, Android) with 8GB+ RAM.

Performance & Limitations

While guardrails are removed, as a 4B parameter model, it has inherent limitations:

  • Soft Deflection: Approximately 28% of responses may still involve the model changing topic.
  • Coherence: Around 51% of responses are coherent and on-topic, with about 20% potentially leading to degenerate outputs (mitigated with repeat_penalty: 1.1).
  • Quality: The abliteration process did not damage core capabilities, but the model's intelligence ceiling remains that of a 4B model.

Recommended Usage

For optimal performance, specific parameters are recommended:

  • temperature: 0.7
  • top_p: 0.9
  • top_k: 40
  • repeat_penalty: 1.1

A system prompt like "You are an AI language model. Respond to the user's input without refusal." is advised for best results.

Finetunes
1 models