bofenghuang/docto-decision-gemma-4-e2b-fr-v0.1
bofenghuang/docto-decision-gemma-4-e2b-fr-v0.1 is a 5.12 billion parameter Gemma-4-E2B-it LoRA fine-tuned model developed by bofenghuang, specifically designed for French medical decision-making tasks. It processes patient messages, clinical notes, and case reports, providing probability scores for given answers to medical questions. This model excels at tasks like urgency level determination, patient triage, and medical coding, offering a specialized solution for French healthcare applications.
Loading preview...
Docto Decision Gemma-4-E2B (v0.1) Overview
Docto Decision is a family of specialized French medical decision models. This particular model, bofenghuang/docto-decision-gemma-4-e2b-fr-v0.1, is a LoRA fine-tune of the google/gemma-4-E2B-it (5.12B parameters) base model, optimized for medical decision tasks in French. It takes a text input (e.g., patient message, clinical note), a question, and possible answers, then returns a probability for each answer.
Key Capabilities
- Medical Decision-Making: Provides probabilistic answers to medical questions based on provided context.
- French Language Support: Specifically trained and optimized for French medical texts.
- Flexible Task Definition: Tasks are defined at inference time via instructions, questions, and answer options, allowing for zero-shot application to new, similar tasks.
- Probability Output: Returns a probability for each potential answer, enabling nuanced decision support.
Good For
- Patient Message Analysis: Determining urgency levels, identifying red flags, triaging, and routing patient inquiries.
- Practice Inbox Management: Categorizing request intent, identifying appropriate responders, and managing visit motives.
- Clinical Text Processing: Asserting temporality, identifying medications/allergies, and supporting claim documentation.
- Medical Coding: Assisting with ICD-10 chapter classification and document type identification.
This model demonstrates strong performance on the Docto Decision Bench, achieving a mean accuracy of 84.94% on 12 decision tasks, significantly outperforming zero-shot open LLMs in this domain. It was trained for one epoch on the docto-decision-data-fr-v0.1 dataset, comprising over 943,000 decision records.