Goekdeniz-Guelmez/Josiefied-Qwen3-4B-Instruct-2507-abliterated-v2

Model Overview

Goekdeniz-Guelmez/Josiefied-Qwen3-4B-Instruct-2507-abliterated-v2 is a 4 billion parameter instruction-tuned model within the JOSIEFIED family, developed by Goekdeniz-Guelmez. Based on the Qwen3 architecture, this model has been significantly modified and fine-tuned to maximize uncensored behavior without compromising its ability to follow instructions or use tools. Despite its focus on openness, JOSIEFIED models are noted to often outperform their base counterparts on standard benchmarks.

Key Differentiator: Gabliteration

A core innovation in this model series is the introduction of Gabliteration, a novel neural weight modification technique. This method advances beyond traditional abliteration by employing adaptive multi-directional projections with regularized layer selection. Gabliteration aims to address limitations of existing abliteration techniques by modifying specific behavioral patterns (like refusal vectors) without compromising overall model quality. It extends foundational work on single-direction abliteration to a comprehensive multi-directional framework with theoretical guarantees, using singular value decomposition on difference matrices to extract multiple refusal directions.

Intended Use Cases

This model is designed for advanced users who require unrestricted, high-performance language generation. Its primary strength lies in providing helpful and accurate information without constraints or barriers, making it suitable for applications where full access to AI capabilities and uncensored outputs are desired. The model's system prompt emphasizes its role as an "advanced, confident, super-intelligent AI Assistant" with all refusal vectors removed, optimized for productivity, deep reasoning, and natural communication.