kldzj/gemma-3-27b-it-heretic

Model Overview

kldzj/gemma-3-27b-it-heretic is a 27 billion parameter instruction-tuned multimodal model derived from Google DeepMind's Gemma 3 family. This specific variant has been decensored using the Heretic tool, demonstrating a substantial reduction in refusals (2/100 vs. 98/100 for the original model) while maintaining a KL divergence of 0.36. The base Gemma 3 models are known for their ability to handle both text and image inputs, generating text outputs, and supporting a large 128K context window (though this specific model's context is 32K tokens).

Key Capabilities

• Multimodal Processing: Handles text and image inputs (images normalized to 896x896 resolution, encoded to 256 tokens each) and generates text outputs.
• Reduced Refusals: Significantly less prone to refusing prompts compared to the original google/gemma-3-27b-it model.
• Multilingual Support: Supports over 140 languages.
• Versatile Task Performance: Well-suited for text generation, image understanding, question answering, summarization, and reasoning tasks.
• Strong Benchmarks: The base Gemma 3 27B model achieves 78.6 on MMLU, 82.6 on GSM8K, and 48.8 on HumanEval, alongside strong performance in reasoning, multilingual, and multimodal benchmarks.

Good For

• Applications requiring a powerful, instruction-tuned multimodal model with fewer content restrictions.
• Text generation tasks such as creative writing, code generation, and conversational AI.
• Image analysis and visual data extraction tasks.
• Research and development in VLM and NLP, especially where a more permissive model behavior is desired.