laladream/gemma-4-12B-it-uncensored

Overview

This model, laladream/gemma-4-12B-it-uncensored, is a 12 billion parameter instruction-tuned language model derived from Google's gemma-4-12B-it. Its primary distinction is the removal of refusal behaviors, making it respond more directly to a wider range of prompts.

Key Capabilities & Features

• Significantly Reduced Refusals: Achieves a refusal rate of approximately 2% on a cross-dataset validation of 686 prompts, with an effective rate near 0% after auditing for false positives. This is a substantial reduction from the baseline model's 99% refusal rate on certain datasets.
• Quality Preservation: The modification process, using norm-preserving biprojected abliteration, ensures no degradation in response coherence or quality.
• Advanced Abliteration Method: Employs a novel technique that decomposes weight rows, projects out refusal directions from the direction component only, and recombines with original magnitude, guaranteeing weight magnitude preservation. This differs from standard projection methods by using per-layer refusal directions and a deterministic single-pass approach.
• Unified Architecture Support: Specifically adapted for the Gemma4Unified architecture, focusing abliteration on the upper decoder layers (L15-47) where refusal signals are concentrated.

Use Cases

This model is particularly well-suited for:

• Research and Development: Exploring model behaviors without inherent refusal mechanisms.
• Applications Requiring Direct Responses: Scenarios where typical safety filters might hinder desired outputs, provided ethical considerations are managed by the user.
• Comparative Analysis: Studying the impact of refusal mechanisms on model performance and utility.