Model Overview

This model, rahul7star/gemma-3-12b-it-heretic, is a modified version of Google's Gemma 3 12B instruction-tuned multimodal language model. It has been 'decensored' using the Heretic v1.0.0 tool, resulting in a substantial reduction in refusal rates (3/100 compared to 97/100 for the original model), as indicated by KL divergence metrics.

Key Capabilities

• Multimodal Input: Handles both text and image inputs, with images normalized to 896x896 resolution and encoded to 256 tokens each.
• Large Context Window: Features a 128K token total input context, enabling processing of extensive information.
• Multilingual Support: Supports over 140 languages.
• Reduced Refusals: Demonstrates significantly fewer refusals in responses compared to the base Gemma 3 12B IT model.
• Diverse Task Performance: Well-suited for a variety of tasks including question answering, summarization, reasoning, and image analysis.

Good For

• Applications requiring less restrictive content generation: Ideal for use cases where the original model's safety filters might be overly cautious.
• Text Generation: Creative writing, code generation, marketing copy, and email drafts.
• Conversational AI: Powering chatbots and virtual assistants.
• Image Understanding: Extracting, interpreting, and summarizing visual data.
• Research and Development: Serving as a foundation for VLM and NLP research, especially for exploring model behavior with reduced inherent censorship.