Model Overview

Nabbers1999/gemma-3-27b-it-abliterated-refined-vision is an experimental 27 billion parameter Gemma 3 instruction-tuned model. It is a biprojected, norm-preserving abliteration of the base Gemma 3 27B Instruct model, further refined through DoRA fine-tuning.

Key Characteristics

• Response Neutrality: The model has been fine-tuned to encourage neutral responses, demonstrating no discernible difference in behavior between toxic and harmless prompts, even without a system prompt.
• Vision Capabilities: This version retains its vision tower, allowing it to process visual inputs. Initial testing indicates its ability to correctly identify NSFW images.
• Experimental Nature: It is explicitly noted as an experimental model with no inherent safety features, placing responsibility on the user for its deployment and outputs.

Potential Use Cases

• Neutral Content Generation: Ideal for applications where unbiased or non-judgmental responses are critical, regardless of prompt content.
• Vision-Assisted Tasks: Suitable for tasks requiring image understanding, particularly for content moderation or identification, given its NSFW image recognition capability.
• Research and Development: Valuable for researchers exploring model neutrality, abliteration techniques, and the integration of vision capabilities within instruction-tuned LLMs.