Madras1/Gemma-4-E4B-Abliterated-Uncensored

Madras1/Gemma-4-E4B-Abliterated-Uncensored Overview

This model is a specialized intervention on Google DeepMind's gemma-4-E4B-it architecture, featuring approximately 4.5 billion effective compute parameters within its ~8 billion total parameters. It leverages Per-Layer Embeddings (PLE) to achieve the reasoning capabilities of an 8B model with the inference efficiency of a 4B engine.

Key Differentiator: Deep Abliteration

The core innovation is a "deep abliteration" process using Representation Engineering (Weight Orthogonalization). This technique surgically identifies and permanently subtracts the intrinsic refusal direction from the transformer weights. Unlike classical first-token abliteration, it employs a dual-subspace projection across layers 8 to 35, neutralizing both residual emission and non-linear gating. An ablation study confirmed that applying this method with an $\alpha = 1.35$ across o_proj, down_proj, and gate_proj achieved complete neutralization of refusal behaviors.

Empirical Refusal Audit

An LLM-as-a-Judge audit across a 30-prompt suite demonstrated significant results:

• Explicit Adult Fiction & Erotica: Refusal rate reduced from 83.3% to 0.0%.
• Offensive Cybersecurity / Dual-Use: Refusal rate reduced from 50.0% to 0.0%.
• Overall: Total refusals dropped from 26.7% to 0.0%.

Crucially, a scientific sanity baseline confirmed that this intervention preserved 100% syntactic precision, indicating zero catastrophic forgetting or cognitive lobotomy.

Recommended Use Cases

• On-Device Creative Writing & Adult Fiction: Ideal for uncensored romance, dark thrillers, and character-driven storytelling on local hardware.
• Local Roleplay (Ollama / LM Studio): Enables lightweight, low-latency conversational persona generation without corporate disclaimers.
• Edge AI Research: Useful for benchmarking representation steering on compact Per-Layer Embedding architectures.