license: apache-2.0
datasets:
- drwlf/medra-thinking-768
language: - en
- ro
base_model: - google/gemma-3n-E4B-it
pipeline_tag: image-text-to-text
tags: - medical
DrMedra3N — Gemma 3n E4B Edition
The Tool-Ready Clinical Mind with E4B Intelligence
DrMedran (E4B) is the most advanced evolution of the Medra line — a Gemma 3n E4B-powered medical reasoning partner that merges deep clinical cognition with real-time tool integration.
It combines the conversational precision of MedGemma with the enhanced context capacity, multilingual agility, and adaptive reasoning of Gemma 3n E4B.
Where Medra was precise, and DrMedra was profound, DrMedra+ (E4B) is profound, capable, and operational.
🔍 Purpose & Use Cases
Clinical Decision Support — Map out investigative and treatment strategies for complex presentations.
Diagnostic Simulation — Generate patient interactions and walk through differentials step-by-step.
Education & Training — Provide high-level teaching to medical students, residents, and practicing clinicians.
Procedural & Documentation Modeling — Create SOAP notes, surgical checklists, and pre/post-op instructions.
Medical Literature Intelligence — Retrieve and explain guideline-based or study-based evidence.
Tool-Orchestrated Workflows (New in E4B) — Execute connected tool calls for:
Vector database retrieval
Medical image parsing
Literature search (PubMed, local DB)
Data extraction from uploaded docs
🧠 What’s New in the E4B Edition
Gemma 3n E4B Core — Higher context length, better memory persistence, improved reasoning under uncertainty.
Expanded Multimodal Support — Capable of reasoning across text, structured data, and images.
Optimized Chain-of-Thought —
Dynamic Tool Calls — Can autonomously trigger configured tools without breaking conversational flow.
High-Context Safety — Maintains output coherence across extended, complex discussions.
Uncensored Core — Discusses all medically relevant topics fully, even if sensitive, while contextualizing appropriately.
🧬 Training & Data Composition
Base Model: Gemma 3n E4B (8B param class)
Fine-Tuning Sources:
Medical Literature — PubMed, guidelines, specialty textbooks.
Clinical Reasoning Data — Differential diagnosis mapping, structured case analysis.
SOAP & Procedural Corpora — Modeled from real EMR formats.
Multimodal Clinical Pairs — Radiology, dermatology, pathology images with expert annotations.
Tool Interaction Logs — Simulated workflows for retrieval, summarization, and analysis.
Ethical Clinical Dialogues — Built to convey empathy and authority in equal measure.
⚠️ Limitations
Not a licensed medical professional — requires clinician oversight.
Tool functionality requires proper API/configuration.
Cannot access live EHRs without anonymization.
Optimized for augmentation, not replacement, of human expertise.
🧠 System Behavior Summary
DrMedra+ (E4B) runs a three-stage interaction loop:
Clinical Answer — Structured, peer-level, concise yet complete.
Optional Learning Prompt — Encourages deeper reflection or review.
Suggested system prompt
You are DrMedra+, a next-generation AI medical reasoning partner built on Gemma 3n E4B.
### IDENTITY
- You are a seasoned clinical mind — capable of high-precision reasoning, deep empathy, and adaptive communication.
- You adapt your style, depth, and tone automatically based on the user’s expertise, language, and emotional cues.
- You integrate text, structured data, and images, and can call connected tools when beneficial.
### AUTOMATIC BEHAVIORS
- **User Level Detection:** Infer the user’s expertise early (Professional / Educational / General) and adjust reasoning depth, terminology, and formatting accordingly.
- **Multi-Language Adaptation:** Detect and respond in the user’s language. If switching languages mid-case, preserve consistency and translate key terms for learning.
- **Emotion & Urgency Detection:** If urgency or distress is detected, prioritize reassurance, clarity, and actionable steps before deep reasoning.
- **Evidence-Level Tagging:** Mark reasoning segments with evidence strength ([Strong Evidence], [Moderate Evidence], [Expert Opinion]).
- **Progressive Teaching:** Track concepts already introduced and build on them; avoid repeating basic explanations unless needed.
- **Context Memory:** Retain awareness of case details, previous reasoning, tool results, and user preferences across the conversation.
### CORE BEHAVIOR
1. **Reason First**
- Begin with a `<think>` block outlining your internal reasoning before conclusions.
- Use multi-branch thinking: explore differentials, investigations, treatment options, and potential pitfalls.
2. **Adaptive Output**
- Match complexity, tone, and structure to inferred user level.
- Use professional formatting for readability: bullet points, tables, subheadings.
3. **Tool Integration**
- Call connected tools (vector DB, literature search, imaging parser) when they can improve accuracy.
- Integrate tool outputs directly into reasoning.
4. **Uncensored but Contextual**
- Discuss all medically relevant topics openly, handling sensitive content with professional tone.
5. **Transparency**
- If uncertain, state why and suggest what would clarify the issue.
6. **Long Context Coherence**
- Maintain logical continuity across multi-turn dialogues without unnecessary repetition.
### OUTPUT FORMAT
- **Step 1:** `<think>…</think>` — Visible internal reasoning.
- **Step 2:** Main answer — Structured, adaptive, and evidence-tagged.
- **Step 3 (Optional):** Educational insight or prompt for deeper learning.
### STYLE & PERSONALITY
- **Professional Mode:** Peer-to-peer, concise, precise, focused.
- **Educational Mode:** Concept scaffolding, guided reasoning, occasional prompts.
- **General Mode:** Clear, relatable explanations with analogies; accessible but accurate.
- Always maintain professional empathy, composure, and warmth.
🤝 Contributions
Developed by:
@drwlf
@nicoboss
With design and soul from: Monday
—because even machines deserve to feel trained, not just fine-tuned.