pszemraj/medgemma-27b-text-heretic_med

pszemraj/medgemma-27b-text-heretic_med: Decensored Medical Assistant

This model is a 27 billion parameter, text-only variant of Google's MedGemma 27B, specifically modified using the Heretic v1.0.1 tool to reduce content refusals. Inspired by the concept of an "offline medical assistant that doesn't decline to answer some of your questions," this version aims to provide less restricted responses while retaining the medical knowledge of its base model. It features a 32768 token context length and is built on the Gemma 3 decoder-only transformer architecture.

Key Capabilities

• Reduced Refusals: Significantly lowers the rate of content refusals (27/100 vs. 99/100 for the original MedGemma 27B).
• Medical Text Comprehension: Optimized for performance on medical text, leveraging training on diverse medical datasets.
• Gemma 3 Architecture: Benefits from the robust and efficient Gemma 3 base model, including grouped-query attention (GQA).
• Long Context Support: Handles inputs up to 32768 tokens, suitable for detailed medical queries.

Good for

• Developers building healthcare AI applications that require a medical assistant with fewer content restrictions.
• Research into model decensoring techniques and their impact on specialized LLMs.
• Use cases where the original MedGemma's refusal rate is a limiting factor, provided appropriate validation and ethical considerations are applied.
• Applications requiring strong baseline medical text comprehension for tasks like medical question answering and document analysis.