bofenghuang/docto-decision-qwen3.5-2b-fr-v0.1

Docto Decision Qwen3.5-2B (v0.1) Overview

This model is a specialized French medical decision model, fine-tuned from Qwen3.5-2B (2.27B parameters) using LoRA. It is designed to take a medical text (e.g., patient message, clinical note), a question, and possible answers, then return a probability for each answer. The model is optimized to respond with a single token, representing an option letter or "oui" / "non", with probabilities derived from that token.

Key Capabilities

• Medical Decision Making: Provides probabilistic answers for various medical scenarios.
• French Language Specialization: Specifically trained and optimized for French medical texts.
• Flexible Task Definition: Tasks are defined at inference time via instructions, questions, and answers, allowing for zero-shot application to new tasks.
• High Accuracy: Achieves 84.91% mean accuracy on the Docto Decision Bench, significantly outperforming its base Qwen3.5-2B model (60.12%) and competitive with larger frontier LLMs like Gemini-2.5-Flash.
• Low Latency: Demonstrates a median latency of 36 ms, making it suitable for real-time applications.

Good For

• Patient Message Triage: Assessing urgency levels, identifying red flags, and routing patients to appropriate specialties.
• Practice Inbox Management: Determining request intent, assigning tasks, and managing reminder replies.
• Clinical Text Analysis: Asserting temporality, identifying medications and allergies, and supporting claim documentation.
• Coding and Document Classification: Categorizing ICD-10 chapters, document types, and article specialties.
• Assistant Safety: Implementing input guardrails, ensuring answer compliance, and checking for personal identifiers in medical contexts.