pszemraj/medgemma-4b-it-heretic
The pszemraj/medgemma-4b-it-heretic is a 4.3 billion parameter instruction-tuned multimodal language model, derived from Google's MedGemma-4B-IT, with a 32768 token context length. This variant has been decensored using the Heretic v1.0.1 tool, significantly reducing refusals compared to the original model. It is optimized for medical text and image comprehension, excelling in tasks like medical image classification, visual question answering, and text-based medical knowledge, making it suitable for healthcare AI applications requiring less restrictive responses.
Loading preview...
pszemraj/medgemma-4b-it-heretic: Decensored Medical LLM
This model is a 4.3 billion parameter instruction-tuned variant of Google's MedGemma-4B-IT, specifically modified using the Heretic v1.0.1 tool to reduce content refusals. While the original MedGemma-4B-IT exhibited 98/100 refusals in testing, this 'heretic' version demonstrates a significantly lower refusal rate of 18/100, offering a less restricted medical assistant experience.
Key Capabilities
- Multimodal Medical Comprehension: Processes both medical text and images (e.g., chest X-rays, dermatology, histopathology, ophthalmology).
- Reduced Refusals: Engineered to provide answers to a broader range of medical questions that the original model might decline.
- Medical Reasoning: Optimized for inference-time computation on medical reasoning tasks.
- Performance: Outperforms the base Gemma 3 4B model across various medical image classification, visual question answering, and text-only medical benchmarks.
- Long Context: Supports a context length of at least 128K tokens.
Good For
- Developers building healthcare-based AI applications requiring a multimodal model with fewer content restrictions.
- Medical text generation and image analysis where a less censored response is desired.
- Experimentation with medical AI models that offer more direct answers to sensitive queries.
Limitations
It is crucial to note that this model is not intended for direct clinical diagnosis or treatment recommendations without appropriate validation and adaptation. All outputs should be independently verified. The decensoring process may alter the safety profile compared to the original, highly-filtered MedGemma model.