Model Overview

This model, genevera/medgemma-27b-it-heretic, is a 27 billion parameter instruction-tuned variant of Google's MedGemma-27B-IT, built upon the Gemma 3 architecture. It has been decensored using the Heretic v1.2.0 tool, resulting in a significantly lower refusal rate (6/100) compared to the original model (99/100) while maintaining a low KL divergence of 0.0536. The model supports a substantial context length of 32768 tokens.

Key Capabilities

• Medical Text Comprehension: Optimized for medical text, question-answering, and reasoning tasks, outperforming base Gemma models on benchmarks like MedQA, MedMCQA, and MMLU Med.
• Reduced Refusals: Demonstrates a substantially lower rate of content refusals, making it more permissive for a wider range of medical inquiries.
• Gemma 3 Foundation: Leverages the robust decoder-only transformer architecture of Gemma 3.

Good for

• Healthcare AI Application Development: Serves as a strong starting point for developers building applications that require medical text generation and reasoning.
• Research in Medical LLMs: Useful for researchers exploring the impact of decensoring on medical language models and their performance characteristics.
• Text-Only Medical Use Cases: Particularly well-suited for applications focused solely on medical text, as this variant is text-only, unlike its multimodal counterparts.