Update: Add Semantic RL Mode (NurseEmbed Integration)

.gitattributes

@@ -33,5 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+viz/semantic_clusters.png filter=lfs diff=lfs merge=lfs -text
 *.png filter=lfs diff=lfs merge=lfs -text
 *.jsonl filter=lfs diff=lfs merge=lfs -text

.github/workflows/ci.yml

@@ -0,0 +1,78 @@
+name: CI - NurseSim-RL
+
+on:
+  push:
+    branches: [ main, develop ]
+  pull_request:
+    branches: [ main ]
+
+jobs:
+  test-environment:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ['3.11']
+    
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+    
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v4
+      with:
+        python-version: ${{ matrix.python-version }}
+    
+    - name: Cache pip dependencies
+      uses: actions/cache@v3
+      with:
+        path: ~/.cache/pip
+        key: ${{ runner.os }}-pip-${{ hashFiles('requirements.txt') }}
+        restore-keys: |
+          ${{ runner.os }}-pip-
+    
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install -r requirements.txt
+        pip install pytest pytest-cov flake8
+    
+    - name: Lint with flake8
+      run: |
+        # Stop the build if there are Python syntax errors or undefined names
+        flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
+        # Exit-zero treats all errors as warnings
+        flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics
+    
+    - name: Test environment registration
+      run: |
+        python -c "import gymnasium as gym; import nursesim_rl; env = gym.make('NurseSim-Triage-v0'); print('✅ Environment registered successfully')"
+    
+    - name: Run environment test
+      run: |
+        python test_env.py
+    
+    - name: Validate dataset generation
+      run: |
+        python generate_dataset.py --num_scenarios 10 --output_file data/test_dataset.json
+        echo "✅ Dataset generation successful"
+    
+    - name: Test A2A Protocol Compliance
+      run: |
+        python tests/test_a2a_compliance.py
+    
+    - name: Test Dual-Mode Configuration
+      run: |
+        chmod +x tests/test_dual_mode.sh
+        chmod +x run.sh
+        bash tests/test_dual_mode.sh
+    
+    - name: Validate Agent Card Schema
+      run: |
+        python -c "
+        import json
+        from pathlib import Path
+        card = json.loads(Path('.well-known/agent-card.json').read_text())
+        assert card['protocol'] == 'a2a/v1.0', 'Invalid protocol version'
+        assert 'capabilities' in card, 'Missing capabilities'
+        print('✅ Agent card schema valid')
+        "

.github/workflows/sync-hf-space.yml

@@ -0,0 +1,57 @@
+name: Sync to Hugging Face Space
+
+on:
+  push:
+    branches: [ main ]
+    paths:
+      - 'app.py'
+      - 'requirements.txt'
+      - 'README.md'
+      - 'Dockerfile'
+      - 'run.sh'
+      - 'agent_main.py'
+      - '.well-known/**'
+      - 'nursesim_rl/**'
+
+jobs:
+  sync-to-hub:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+      with:
+        fetch-depth: 0
+        lfs: true
+    
+    - name: Push to Hugging Face Space
+      env:
+        HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      run: |
+        git config --global user.email "actions@github.com"
+        git config --global user.name "GitHub Actions"
+        
+        # Clone the HF Space repo
+        git clone https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo hf-space
+        
+        # Copy updated files
+        cp app.py hf-space/app.py
+        cp requirements.txt hf-space/requirements.txt
+        cp max-requirements.txt hf-space/max-requirements.txt 2>/dev/null || true
+        cp README.md hf-space/README.md
+        cp Dockerfile hf-space/Dockerfile
+        cp run.sh hf-space/run.sh
+        cp agent_main.py hf-space/agent_main.py
+        
+        # Copy directories
+        rm -rf hf-space/.well-known hf-space/nursesim_rl
+        cp -r .well-known hf-space/
+        cp -r nursesim_rl hf-space/
+        
+        # Push to HF Space
+        cd hf-space
+        git add .
+        git status
+        git commit -m "🚀 Auto-sync from GitHub: ${GITHUB_SHA::7}" || echo "No changes to commit"
+        git push https://NurseCitizenDeveloper:$HF_TOKEN@huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo main
+        
+        echo "✅ Successfully synced to Hugging Face Space!"

.gitignore

@@ -0,0 +1,51 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Project specific
+outputs/
+*.log
+wandb/

.well-known/agent-card.json

@@ -1,118 +1,118 @@
-{
-    "name": "NurseSim-Triage",
-    "version": "1.0.0",
-    "description": "An AI agent fine-tuned for clinical triage using the Manchester Triage System (MTS). Provides emergency department triage assessments based on patient complaints and vital signs.",
-    "author": "NurseCitizenDeveloper (ClinyQAi)",
-    "protocol": "a2a/v1.0",
-    "capabilities": {
-        "tasks": [
-            "clinical_triage",
-            "patient_assessment",
-            "vital_signs_evaluation"
-        ],
-        "input_schema": {
-            "type": "object",
-            "properties": {
-                "complaint": {
-                    "type": "string",
-                    "description": "Patient's chief complaint or presenting symptoms"
-                },
-                "vitals": {
-                    "type": "object",
-                    "properties": {
-                        "heart_rate": {
-                            "type": "integer",
-                            "description": "Heart rate in beats per minute"
-                        },
-                        "blood_pressure": {
-                            "type": "string",
-                            "description": "Blood pressure in format '120/80'"
-                        },
-                        "spo2": {
-                            "type": "integer",
-                            "description": "Oxygen saturation percentage"
-                        },
-                        "temperature": {
-                            "type": "number",
-                            "description": "Body temperature in Celsius"
-                        }
-                    },
-                    "required": [
-                        "heart_rate",
-                        "blood_pressure",
-                        "spo2",
-                        "temperature"
-                    ]
-                }
-            },
-            "required": [
-                "complaint",
-                "vitals"
-            ]
-        },
-        "output_schema": {
-            "type": "object",
-            "properties": {
-                "triage_category": {
-                    "type": "string",
-                    "enum": [
-                        "Immediate",
-                        "Very Urgent",
-                        "Urgent",
-                        "Standard",
-                        "Non-Urgent"
-                    ],
-                    "description": "Manchester Triage System category"
-                },
-                "assessment": {
-                    "type": "string",
-                    "description": "Detailed clinical assessment and rationale"
-                },
-                "recommended_action": {
-                    "type": "string",
-                    "description": "Recommended next steps for patient care"
-                }
-            },
-            "required": [
-                "triage_category",
-                "assessment"
-            ]
-        }
-    },
-    "model": {
-        "type": "llm",
-        "base_model": "meta-llama/Llama-3.2-3B-Instruct",
-        "adapter": "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B",
-        "training_method": "LoRA (Low-Rank Adaptation)",
-        "framework": "Hugging Face Transformers + PEFT"
-    },
-    "links": {
-        "repository": "https://github.com/ClinyQAi/NurseSim-RL",
-        "model_card": "https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B",
-        "demo": "https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo",
-        "blog_post": "https://huggingface.co/blog/NurseCitizenDeveloper/nursesim-rl-training-ai-agents-clinical-triage",
-        "agentbeats_profile": "https://agentbeats.dev/ClinyQAi/nursesim-triage"
-    },
-    "requirements": {
-        "gpu": true,
-        "gpu_memory_gb": 8,
-        "python_version": ">=3.9",
-        "environment_variables": [
-            {
-                "name": "HF_TOKEN",
-                "required": true,
-                "description": "Hugging Face API token for accessing gated models"
-            }
-        ]
-    },
-    "license": "Llama 3.2 Community License",
-    "tags": [
-        "healthcare",
-        "triage",
-        "nursing",
-        "clinical-assessment",
-        "medical-education",
-        "llama-3.2",
-        "openenv-challenge"
-    ]
+{
+    "name": "NurseSim-Triage",
+    "version": "1.0.0",
+    "description": "An AI agent fine-tuned for clinical triage using the Manchester Triage System (MTS). Provides emergency department triage assessments based on patient complaints and vital signs.",
+    "author": "NurseCitizenDeveloper (ClinyQAi)",
+    "protocol": "a2a/v1.0",
+    "capabilities": {
+        "tasks": [
+            "clinical_triage",
+            "patient_assessment",
+            "vital_signs_evaluation"
+        ],
+        "input_schema": {
+            "type": "object",
+            "properties": {
+                "complaint": {
+                    "type": "string",
+                    "description": "Patient's chief complaint or presenting symptoms"
+                },
+                "vitals": {
+                    "type": "object",
+                    "properties": {
+                        "heart_rate": {
+                            "type": "integer",
+                            "description": "Heart rate in beats per minute"
+                        },
+                        "blood_pressure": {
+                            "type": "string",
+                            "description": "Blood pressure in format '120/80'"
+                        },
+                        "spo2": {
+                            "type": "integer",
+                            "description": "Oxygen saturation percentage"
+                        },
+                        "temperature": {
+                            "type": "number",
+                            "description": "Body temperature in Celsius"
+                        }
+                    },
+                    "required": [
+                        "heart_rate",
+                        "blood_pressure",
+                        "spo2",
+                        "temperature"
+                    ]
+                }
+            },
+            "required": [
+                "complaint",
+                "vitals"
+            ]
+        },
+        "output_schema": {
+            "type": "object",
+            "properties": {
+                "triage_category": {
+                    "type": "string",
+                    "enum": [
+                        "Immediate",
+                        "Very Urgent",
+                        "Urgent",
+                        "Standard",
+                        "Non-Urgent"
+                    ],
+                    "description": "Manchester Triage System category"
+                },
+                "assessment": {
+                    "type": "string",
+                    "description": "Detailed clinical assessment and rationale"
+                },
+                "recommended_action": {
+                    "type": "string",
+                    "description": "Recommended next steps for patient care"
+                }
+            },
+            "required": [
+                "triage_category",
+                "assessment"
+            ]
+        }
+    },
+    "model": {
+        "type": "llm",
+        "base_model": "meta-llama/Llama-3.2-3B-Instruct",
+        "adapter": "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B",
+        "training_method": "LoRA (Low-Rank Adaptation)",
+        "framework": "Hugging Face Transformers + PEFT"
+    },
+    "links": {
+        "repository": "https://github.com/ClinyQAi/NurseSim-RL",
+        "model_card": "https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B",
+        "demo": "https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo",
+        "blog_post": "https://huggingface.co/blog/NurseCitizenDeveloper/nursesim-rl-training-ai-agents-clinical-triage",
+        "agentbeats_profile": "https://agentbeats.dev/ClinyQAi/nursesim-triage"
+    },
+    "requirements": {
+        "gpu": true,
+        "gpu_memory_gb": 8,
+        "python_version": ">=3.9",
+        "environment_variables": [
+            {
+                "name": "HF_TOKEN",
+                "required": true,
+                "description": "Hugging Face API token for accessing gated models"
+            }
+        ]
+    },
+    "license": "Llama 3.2 Community License",
+    "tags": [
+        "healthcare",
+        "triage",
+        "nursing",
+        "clinical-assessment",
+        "medical-education",
+        "llama-3.2",
+        "openenv-challenge"
+    ]
 }

Dockerfile

@@ -1,36 +1,36 @@
-# Use Python 3.11 base image (required for agentbeats SDK)
-FROM python:3.11-slim
-
-# Set working directory
-WORKDIR /app
-
-# Install system dependencies
-RUN apt-get update && apt-get install -y \
-    git \
-    build-essential \
-    python3-dev \
-    && rm -rf /var/lib/apt/lists/*
-
-# Copy requirements first for caching
-COPY requirements.txt .
-
-# Install Python dependencies
-RUN pip install --no-cache-dir -r requirements.txt
-
-# Copy the rest of the application
-COPY . .
-
-# Make run.sh executable
-RUN chmod +x run.sh
-
-# Expose ports
-# 7860: Gradio demo
-# 8080: AgentBeats A2A controller
-EXPOSE 7860 8080
-
-# Set environment variables
-ENV PYTHONUNBUFFERED=1
-ENV AGENT_MODE=a2a
-
-# Run the application via launcher script
-CMD ["./run.sh"]
+# Use Python 3.11 base image (required for agentbeats SDK)
+FROM python:3.11-slim
+
+# Set working directory
+WORKDIR /app
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    git \
+    build-essential \
+    python3-dev \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy requirements first for caching
+COPY requirements.txt .
+
+# Install Python dependencies
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy the rest of the application
+COPY . .
+
+# Make run.sh executable
+RUN chmod +x run.sh
+
+# Expose ports
+# 7860: Gradio demo
+# 8080: AgentBeats A2A controller
+EXPOSE 7860 8080
+
+# Set environment variables
+ENV PYTHONUNBUFFERED=1
+ENV AGENT_MODE=a2a
+
+# Run the application via launcher script
+CMD ["./run.sh"]

HF_BLOG_POST.md

@@ -0,0 +1,284 @@
+---
+title: "NurseSim-RL: Training AI Agents for Clinical Triage"
+thumbnail: /blog/assets/nursesim-rl/thumbnail.png
+authors:
+- user: NurseCitizenDeveloper
+tags:
+- reinforcement-learning
+- healthcare
+- openenv
+- llama
+- unsloth
+- clinical-ai
+---
+
+# NurseSim-RL: Training AI Agents for Clinical Triage
+
+**TL;DR:** We built a Gymnasium-compatible RL environment that simulates Emergency Department triage and fine-tuned a Llama 3.2 3B model to master it using Unsloth. The agent achieves expert-level performance in assigning Manchester Triage System categories while maintaining safety-critical decision-making.
+
+🔗 **[Live Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)** | **[GitHub](https://github.com/ClinyQAi/NurseSim-RL)** | **[Model](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)**
+
+---
+
+## The Challenge: OpenEnv 2026
+
+This project was developed for the [OpenEnv Challenge](https://rdi.berkeley.edu/agentx-agentbeats), sponsored by PyTorch, Hugging Face, and Unsloth. The goal? Create innovative RL environments that push the boundaries of agentic AI and contribute them as open-source public goods.
+
+Healthcare seemed like the perfect domain—it's **safety-critical**, **high-stakes**, and requires **complex reasoning**. If we can build agents that make good clinical decisions, we're not just advancing AI research; we're potentially saving lives.
+
+---
+
+## The Problem: A&E Triage is Hard
+
+Every day, Emergency Departments (A&E in the UK, ER in the US) face a critical challenge: **which patient gets seen first?**
+
+Triage nurses use the **Manchester Triage System (MTS)** to categorize patients into 5 priority levels:
+
+| Category | Priority | Target Time | Example |
+|----------|----------|-------------|---------|
+| **1** | Immediate | 0 min | Cardiac arrest, Anaphylaxis |
+| **2** | Very Urgent | 10 min | Chest pain (STEMI), Stroke |
+| **3** | Urgent | 60 min | Abdominal pain, Fractures |
+| **4** | Standard | 120 min | Minor injuries, Viral illness |
+| **5** | Non-Urgent | 240 min | Minor cuts, GP-suitable |
+
+### Why This Matters
+
+A wrong decision has real consequences:
+- **Under-triage** a Category 1 patient → Life-threatening delay
+- **Over-triage** a Category 5 patient → Wasted critical resources
+
+This isn't just a classification problem—it's a **safety-critical resource allocation game**.
+
+---
+
+## The Solution: NurseSim-RL Environment
+
+We built `NurseSim-Triage-v0`, a Gymnasium-compatible environment that models the A&E triage workflow.
+
+### How It Works
+
+**Observation Space:**
+```python
+{
+  "patient_complaint": "Crushing chest pain radiating to left arm",
+  "vitals": {
+    "HR": 110,
+    "BP": "90/60",
+    "SpO2": 94,
+    "Temp": 37.2
+  },
+  "waiting_room": 8,
+  "available_beds": 2
+}
+```
+
+**Action Space:**
+```python
+{
+  "triage_category": 2,  # 1-5 (MTS)
+  "intervention": "send_to_resus"  # Clinical action
+}
+```
+
+**Reward Function:**
+- **+10** for correct triage category
+- **-50** for critical safety failures (e.g., discharging a Cat 1 patient)
+- **-1** per minute of wait time for critical patients
+
+### Dataset Generation
+
+We created a `PatientGenerator` class that produces realistic scenarios:
+- **500 training examples** covering all 5 MTS categories
+- Realistic vital sign variations (e.g., tachycardia in sepsis, hypotension in shock)
+- Distribution mimicking real A&E patient flow (more Cat 3-4 than Cat 1-2)
+
+**Example:**
+```json
+{
+  "instruction": "You are an expert A&E Triage Nurse...",
+  "input": "Patient: 68-year-old male, crushing chest pain...",
+  "output": "CATEGORY 2 (Very Urgent). Rationale: Classic STEMI presentation..."
+}
+```
+
+---
+
+## Training: Llama 3.2 + Unsloth = Magic ✨
+
+We used **Unsloth** to fine-tune `Llama-3.2-3B-Instruct` with 4-bit QLoRA. Why Unsloth? **2x faster training** and **60% less memory**.
+
+### Setup
+
+```python
+from unsloth import FastLanguageModel
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name="unsloth/Llama-3.2-3B-Instruct",
+    max_seq_length=2048,
+    load_in_4bit=True,
+)
+
+model = FastLanguageModel.get_peft_model(
+    model,
+    r=16,
+    lora_alpha=16,
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
+                    "gate_proj", "up_proj", "down_proj"],
+)
+```
+
+### Training Results
+
+The convergence was **stunning**:
+
+| Metric | Value |
+|--------|-------|
+| Initial Loss | 2.8 |
+| Final Loss | **0.08** |
+| Steps | 100 |
+| Epochs | ~6 |
+| Hardware | NVIDIA A100 (Colab) |
+| Time | **15 minutes** |
+
+![Training Loss Curve](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/train_loss.png)
+
+*The training loss dropped from 2.8 to <0.1 in just 100 steps, demonstrating rapid domain adaptation.*
+
+The model went from "guessing" to "expert" in just 100 optimization steps. This rapid domain adaptation shows that **LLMs can learn specialized clinical reasoning with minimal compute**.
+
+### Training Metrics Deep Dive
+
+Below are the complete training metrics from our W&B run:
+
+<details>
+<summary><b>📈 Click to View All Training Charts</b></summary>
+
+#### Loss Progression (Global Steps)
+![Loss by Global Step](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/Train-globalstep.png)
+
+#### Loss Progression (Epochs)
+![Loss by Epoch](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/Train-epoch.png)
+
+#### Gradient Norm Stability
+![Gradient Norm](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/Train-grad_norm.png)
+*Gradient norm stabilized after ~20 steps, indicating healthy convergence.*
+
+#### Learning Rate Schedule
+![Learning Rate](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/train_learningrate.png)
+*Linear decay from 2e-4 to 0 over 100 steps.*
+
+</details>
+
+**Key Observations:**
+- ✅ **No overfitting:** Smooth loss curve without erratic spikes
+- ✅ **Stable gradients:** No exploding/vanishing gradient issues
+- ✅ **Efficient optimization:** Reached convergence well before max_steps
+
+---
+
+## The Agent in Action
+
+We deployed the fine-tuned model to a **Gradio Space** powered by ZeroGPU:
+
+**[🩺 Try the Live Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)**
+
+### Example Inference
+
+**Input:**
+```
+Chief Complaint: "Crushing chest pain and nausea"
+Vitals: HR 110, BP 90/60, SpO2 94%
+```
+
+**Output:**
+```
+CATEGORY 2 (Very Urgent - 10 min target)
+
+Rationale: Classic presentation of acute coronary syndrome (ACS). 
+The crushing chest pain combined with hypotension (BP 90/60) and 
+mild hypoxia (SpO2 94%) indicates significant cardiac compromise.
+
+Recommended Action: Immediate ECG, troponin, aspirin 300mg, IV access.
+Send to Resus for continuous monitoring.
+```
+
+The agent not only assigns the correct category but also **explains its reasoning** and **recommends clinical actions**—behaviors learned purely from the training data.
+
+---
+
+## Technical Deep Dive
+
+### Why Llama 3.2?
+
+1. **Instruction-tuned:** Already aligned for conversational tasks
+2. **Small enough for edge deployment:** 3B parameters = mobile/browser inference
+3. **Meta's clinical pre-training:** Better baseline than general-purpose models
+
+### Why 4-bit QLoRA?
+
+- **Memory:** Fits on consumer GPUs (even T4!)
+- **Speed:** Unsloth's kernel optimizations make it viable
+- **Accuracy:** Minimal degradation vs full fine-tuning for this task
+
+### Reproducibility
+
+Everything is open-source:
+- **Dockerfile:** `docker build -t nursesim . && docker run -p 7860:7860 nursesim`
+- **Colab Notebook:** One-click training replication
+- **GitHub:** Full environment code + tests
+
+---
+
+## Lessons Learned
+
+### What Worked
+
+1. **Synthetic data quality matters more than quantity:** 500 well-crafted examples > 10,000 noisy ones
+2. **Unsloth is a game-changer:** Training went from "weekend project" to "15 minutes"
+3. **Safety constraints are learnable:** The model respects the -50 penalty and rarely under-triages
+
+### What Could Be Better
+
+1. **Real clinical validation:** We need nurses to red-team the system
+2. **Uncertainty quantification:** The model should say "I don't know" when confidence is low
+3. **Multi-modal inputs:** Real triage uses visual cues (patient appearance, distress level)
+
+---
+
+## Impact & Future Work
+
+### Immediate Applications
+
+- **Nursing Education:** Students can practice triage scenarios 24/7
+- **Workforce Augmentation:** AI-assisted triage in low-resource settings
+- **Benchmarking:** Other researchers can use NurseSim-RL to test their agents
+
+### Next Steps
+
+1. **Partner with NHS Trusts** for real-world pilot testing
+2. **Extend to other clinical domains** (radiology, discharge planning)
+3. **Build multi-agent systems** (Triage Nurse + Consultant + Pharmacist)
+
+---
+
+## Try It Yourself
+
+All the code, data, and models are open-source:
+
+- 🎮 **[Live Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)**
+- 💻 **[GitHub Repo](https://github.com/ClinyQAi/NurseSim-RL)**
+- 🤗 **[Model on HF Hub](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)**
+- 📓 **[Training Notebook](https://github.com/ClinyQAi/NurseSim-RL/blob/main/notebooks/NurseSim_RL_Unsloth_Training.ipynb)**
+
+---
+
+## Acknowledgements
+
+- **OpenEnv Challenge** - Berkeley RDI, PyTorch, Hugging Face, Unsloth
+- **Manchester Triage System** - Clinical framework
+- **Unsloth AI** - For making LLM fine-tuning actually enjoyable
+
+---
+
+*Built with ❤️ for the OpenEnv Challenge 2026*

HF_BLOG_POST_CLEAN.md

@@ -0,0 +1,270 @@
+# NurseSim-RL: Training AI Agents for Clinical Triage
+
+**TL;DR:** We built a Gymnasium-compatible RL environment that simulates Emergency Department triage and fine-tuned a Llama 3.2 3B model to master it using Unsloth. The agent achieves expert-level performance in assigning Manchester Triage System categories while maintaining safety-critical decision-making.
+
+🔗 **[Live Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)** | **[GitHub](https://github.com/ClinyQAi/NurseSim-RL)** | **[Model](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)**
+
+---
+
+## The Challenge: OpenEnv 2026
+
+This project was developed for the [OpenEnv Challenge](https://rdi.berkeley.edu/agentx-agentbeats), sponsored by PyTorch, Hugging Face, and Unsloth. The goal? Create innovative RL environments that push the boundaries of agentic AI and contribute them as open-source public goods.
+
+Healthcare seemed like the perfect domain—it's **safety-critical**, **high-stakes**, and requires **complex reasoning**. If we can build agents that make good clinical decisions, we're not just advancing AI research; we're potentially saving lives.
+
+---
+
+## The Problem: A&E Triage is Hard
+
+Every day, Emergency Departments (A&E in the UK, ER in the US) face a critical challenge: **which patient gets seen first?**
+
+Triage nurses use the **Manchester Triage System (MTS)** to categorize patients into 5 priority levels:
+
+| Category | Priority | Target Time | Example |
+|----------|----------|-------------|---------|
+| **1** | Immediate | 0 min | Cardiac arrest, Anaphylaxis |
+| **2** | Very Urgent | 10 min | Chest pain (STEMI), Stroke |
+| **3** | Urgent | 60 min | Abdominal pain, Fractures |
+| **4** | Standard | 120 min | Minor injuries, Viral illness |
+| **5** | Non-Urgent | 240 min | Minor cuts, GP-suitable |
+
+### Why This Matters
+
+A wrong decision has real consequences:
+- **Under-triage** a Category 1 patient → Life-threatening delay
+- **Over-triage** a Category 5 patient → Wasted critical resources
+
+This isn't just a classification problem—it's a **safety-critical resource allocation game**.
+
+---
+
+## The Solution: NurseSim-RL Environment
+
+We built `NurseSim-Triage-v0`, a Gymnasium-compatible environment that models the A&E triage workflow.
+
+### How It Works
+
+**Observation Space:**
+```python
+{
+  "patient_complaint": "Crushing chest pain radiating to left arm",
+  "vitals": {
+    "HR": 110,
+    "BP": "90/60",
+    "SpO2": 94,
+    "Temp": 37.2
+  },
+  "waiting_room": 8,
+  "available_beds": 2
+}
+```
+
+**Action Space:**
+```python
+{
+  "triage_category": 2,  # 1-5 (MTS)
+  "intervention": "send_to_resus"  # Clinical action
+}
+```
+
+**Reward Function:**
+- **+10** for correct triage category
+- **-50** for critical safety failures (e.g., discharging a Cat 1 patient)
+- **-1** per minute of wait time for critical patients
+
+### Dataset Generation
+
+We created a `PatientGenerator` class that produces realistic scenarios:
+- **500 training examples** covering all 5 MTS categories
+- Realistic vital sign variations (e.g., tachycardia in sepsis, hypotension in shock)
+- Distribution mimicking real A&E patient flow (more Cat 3-4 than Cat 1-2)
+
+**Example:**
+```json
+{
+  "instruction": "You are an expert A&E Triage Nurse...",
+  "input": "Patient: 68-year-old male, crushing chest pain...",
+  "output": "CATEGORY 2 (Very Urgent). Rationale: Classic STEMI presentation..."
+}
+```
+
+---
+
+## Training: Llama 3.2 + Unsloth = Magic ✨
+
+We used **Unsloth** to fine-tune `Llama-3.2-3B-Instruct` with 4-bit QLoRA. Why Unsloth? **2x faster training** and **60% less memory**.
+
+### Setup
+
+```python
+from unsloth import FastLanguageModel
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name="unsloth/Llama-3.2-3B-Instruct",
+    max_seq_length=2048,
+    load_in_4bit=True,
+)
+
+model = FastLanguageModel.get_peft_model(
+    model,
+    r=16,
+    lora_alpha=16,
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
+                    "gate_proj", "up_proj", "down_proj"],
+)
+```
+
+### Training Results
+
+The convergence was **stunning**:
+
+| Metric | Value |
+|--------|-------|
+| Initial Loss | 2.8 |
+| Final Loss | **0.08** |
+| Steps | 100 |
+| Epochs | ~6 |
+| Hardware | NVIDIA A100 (Colab) |
+| Time | **15 minutes** |
+
+![Training Loss Curve](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/train_loss.png)
+
+*The training loss dropped from 2.8 to <0.1 in just 100 steps, demonstrating rapid domain adaptation.*
+
+The model went from "guessing" to "expert" in just 100 optimization steps. This rapid domain adaptation shows that **LLMs can learn specialized clinical reasoning with minimal compute**.
+
+### Training Metrics Deep Dive
+
+Below are the complete training metrics from our W&B run:
+
+<details>
+<summary><b>📈 Click to View All Training Charts</b></summary>
+
+#### Loss Progression (Global Steps)
+![Loss by Global Step](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/Train-globalstep.png)
+
+#### Loss Progression (Epochs)
+![Loss by Epoch](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/Train-epoch.png)
+
+#### Gradient Norm Stability
+![Gradient Norm](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/Train-grad_norm.png)
+*Gradient norm stabilized after ~20 steps, indicating healthy convergence.*
+
+#### Learning Rate Schedule
+![Learning Rate](https://raw.githubusercontent.com/ClinyQAi/NurseSim-RL/main/docs/train_learningrate.png)
+*Linear decay from 2e-4 to 0 over 100 steps.*
+
+</details>
+
+**Key Observations:**
+- ✅ **No overfitting:** Smooth loss curve without erratic spikes
+- ✅ **Stable gradients:** No exploding/vanishing gradient issues
+- ✅ **Efficient optimization:** Reached convergence well before max_steps
+
+---
+
+## The Agent in Action
+
+We deployed the fine-tuned model to a **Gradio Space** powered by ZeroGPU:
+
+**[🩺 Try the Live Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)**
+
+### Example Inference
+
+**Input:**
+```
+Chief Complaint: "Crushing chest pain and nausea"
+Vitals: HR 110, BP 90/60, SpO2 94%
+```
+
+**Output:**
+```
+CATEGORY 2 (Very Urgent - 10 min target)
+
+Rationale: Classic presentation of acute coronary syndrome (ACS). 
+The crushing chest pain combined with hypotension (BP 90/60) and 
+mild hypoxia (SpO2 94%) indicates significant cardiac compromise.
+
+Recommended Action: Immediate ECG, troponin, aspirin 300mg, IV access.
+Send to Resus for continuous monitoring.
+```
+
+The agent not only assigns the correct category but also **explains its reasoning** and **recommends clinical actions**—behaviors learned purely from the training data.
+
+---
+
+## Technical Deep Dive
+
+### Why Llama 3.2?
+
+1. **Instruction-tuned:** Already aligned for conversational tasks
+2. **Small enough for edge deployment:** 3B parameters = mobile/browser inference
+3. **Meta's clinical pre-training:** Better baseline than general-purpose models
+
+### Why 4-bit QLoRA?
+
+- **Memory:** Fits on consumer GPUs (even T4!)
+- **Speed:** Unsloth's kernel optimizations make it viable
+- **Accuracy:** Minimal degradation vs full fine-tuning for this task
+
+### Reproducibility
+
+Everything is open-source:
+- **Dockerfile:** `docker build -t nursesim . && docker run -p 7860:7860 nursesim`
+- **Colab Notebook:** One-click training replication
+- **GitHub:** Full environment code + tests
+
+---
+
+## Lessons Learned
+
+### What Worked
+
+1. **Synthetic data quality matters more than quantity:** 500 well-crafted examples > 10,000 noisy ones
+2. **Unsloth is a game-changer:** Training went from "weekend project" to "15 minutes"
+3. **Safety constraints are learnable:** The model respects the -50 penalty and rarely under-triages
+
+### What Could Be Better
+
+1. **Real clinical validation:** We need nurses to red-team the system
+2. **Uncertainty quantification:** The model should say "I don't know" when confidence is low
+3. **Multi-modal inputs:** Real triage uses visual cues (patient appearance, distress level)
+
+---
+
+## Impact & Future Work
+
+### Immediate Applications
+
+- **Nursing Education:** Students can practice triage scenarios 24/7
+- **Workforce Augmentation:** AI-assisted triage in low-resource settings
+- **Benchmarking:** Other researchers can use NurseSim-RL to test their agents
+
+### Next Steps
+
+1. **Partner with NHS Trusts** for real-world pilot testing
+2. **Extend to other clinical domains** (radiology, discharge planning)
+3. **Build multi-agent systems** (Triage Nurse + Consultant + Pharmacist)
+
+---
+
+## Try It Yourself
+
+All the code, data, and models are open-source:
+
+- 🎮 **[Live Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)**
+- 💻 **[GitHub Repo](https://github.com/ClinyQAi/NurseSim-RL)**
+- 🤗 **[Model on HF Hub](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)**
+- 📓 **[Training Notebook](https://github.com/ClinyQAi/NurseSim-RL/blob/main/notebooks/NurseSim_RL_Unsloth_Training.ipynb)**
+
+---
+
+## Acknowledgements
+
+- **OpenEnv Challenge** - Berkeley RDI, PyTorch, Hugging Face, Unsloth
+- **Manchester Triage System** - Clinical framework
+- **Unsloth AI** - For making LLM fine-tuning actually enjoyable
+
+---
+
+*Built with ❤️ for the OpenEnv Challenge 2026*

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 NurseCitizenDeveloper
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

MODEL_CARD.md

@@ -0,0 +1,117 @@
+---
+license: llama3.2
+base_model: unsloth/Llama-3.2-3B-Instruct
+tags:
+- reinforcement-learning
+- OpenEnv
+- medical
+- nursing
+- triage
+- gymnasium
+- unsloth
+- lora
+- trl
+- text-generation-inference
+model-index:
+- name: NurseSim-Triage-Llama-3.2-3B
+  results:
+  - task:
+      type: reinforcement-learning
+      name: Nursing Triage (Manchester Triage System)
+    dataset:
+      name: NurseSim-RL-Synthetic-Triage
+      type: synthetic
+    metrics:
+    - type: mean_reward
+      value: 12.5
+      name: Mean Episode Reward (Correct Triage)
+---
+
+# NurseSim-Triage-Llama-3.2-3B
+
+**A state-of-the-art Reinforcement Learning agent for Emergency Department Triage.**
+
+This model is a fine-tuned version of `Llama-3.2-3B-Instruct` using **Unsloth** and **LoRA**. It was developed as part of the **OpenEnv Challenge** to demonstrate agentic reasoning in complex healthcare environments.
+
+## Model Description
+
+- **Task:** Clinical Triage Decision Support
+- **Environment:** `NurseSim-Triage-v0` (Gymnasium-compatible)
+- **Framework:** Manchester Triage System (MTS)
+- **Fine-tuning Strategy:** Supervised Fine-Tuning (SFT) + RL ready architecture.
+- **Quantization:** 4-bit (bitsandbytes) for efficient execution.
+
+## Intended Use & Clinical Rationale
+
+This model is designed to simulate the decision-making process of a Triage Nurse in an Accident & Emergency (A&E) setting. It evaluates:
+1. **Chief Complaint:** Natural language processing of patient symptoms.
+2. **Vitals:** Quantitative analysis of HR, BP, SpO2, and Temperature.
+3. **Safety:** Mitigation of "under-triaging" critical patients (Cat 1/2).
+
+> [!WARNING]
+> **NOT FOR MEDICAL USE.** This model is a research artifact developed for the OpenEnv Challenge. It should not be used in live clinical environments for patient care.
+
+## Training Details
+
+### Dataset
+Trained on a diverse set of synthetic patient scenarios (n=500) covering:
+- **Category 1 (Immediate):** Cardiac arrest, Anaphylaxis, Major Trauma.
+- **Category 2 (Very Urgent):** Chest pain (STEMI), Stroke, Sepsis.
+- **Category 3-5:** Minor injuries, viral illnesses, and primary care redirects.
+
+### Procedure
+- **Optimizer:** AdamW (8-bit)
+- **Learning Rate:** 2e-4
+- **Rank (r):** 16
+- **Alpha:** 16
+- **Hardware:** Trained on NVIDIA A100 (Google Colab High-RAM).
+- **Time:** ~15 minutes with Unsloth optimization.
+
+## Evaluation & Training Results
+
+### Convergence Overview
+The model showed rapid and stable convergence during its 100-step training run:
+- **Loss Reduction:** Training loss dropped significantly from an initial **2.8** to a terminal value of **<0.1** within approximately 6 epochs.
+- **Gradient Stability:** `grad_norm` stabilized after step 20, indicating a highly compatible dataset for the Llama 3.2 architecture.
+- **Learning Rate:** Used a linear warmup to 2e-4 followed by a linear decay to zero.
+
+### Performance Metrics (Environment: NurseSim-Triage-v0)
+
+| Category | Performance | Outcome |
+|----------|-------------|---------|
+| Loss | ~0.08 | Near-perfect alignment with expert triage decisions. |
+| Steps | 100 | Sufficient for specialized domain adaptation. |
+| Epochs | 6+ | Ensuring deep extraction of MTS patterns. |
+
+## How to use
+
+```python
+from unsloth import FastLanguageModel
+import torch
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B",
+    max_seq_length = 2048,
+    load_in_4bit = True,
+)
+FastLanguageModel.for_inference(model)
+
+# Assessment Prompt
+prompt = """### Instruction:
+You are an expert A&E Triage Nurse. Assess the following patient and provide your triage decision.
+
+### Input:
+Patient presents with crushing central chest pain radiating to left arm. 
+Vitals: HR 110, BP 90/60, SpO2 94%.
+
+### Response:"""
+
+inputs = tokenizer([prompt], return_tensors = "pt").to("cuda")
+outputs = model.generate(**inputs, max_new_tokens = 256)
+tokenizer.batch_decode(outputs)
+```
+
+## Acknowledgements
+- **OpenEnv Team** for the challenge framework.
+- **Unsloth AI** for the 2x faster training tools.
+- **Meta Llama** for the base architecture.

README.md

@@ -1,330 +1,331 @@
----
-title: NurseSim Triage
-emoji: 🏥
-colorFrom: blue
-colorTo: indigo
-sdk: docker
-pinned: false
----
-
-# NurseSim-RL: A Healthcare Agent Environment for Clinical Triage
-
-[![AgentBeats A2A](https://img.shields.io/badge/AgentBeats-A2A%20Enabled-purple)](https://agentbeats.dev/ClinyQAi/nursesim-triage)
-
-[![OpenEnv Challenge](https://img.shields.io/badge/OpenEnv-Challenge%202026-blue)](https://rdi.berkeley.edu/agentx-agentbeats)
-[![Hugging Face Model](https://img.shields.io/badge/🤗-Model-yellow)](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)
-[![W&B Report](https://img.shields.io/badge/W%26B-Report-orange)](https://wandb.ai/mrlincs-nursing-citizen-development/huggingface)
-[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
-
-> **OpenEnv Challenge Entry** | Berkeley RDI AgentX-AgentBeats Competition  
-> A Gymnasium-compatible RL environment for training AI agents to perform clinical triage using the Manchester Triage System (MTS).
-
-![NurseSim Demo](docs/demo.gif)
-
-## 🎯 Overview
-
-**NurseSim-RL** simulates the decision-making process of a Triage Nurse in an Accident & Emergency (A&E) department. The agent must assess patients based on their chief complaint and vital signs, then assign an appropriate triage category (1-5) according to the Manchester Triage System.
-
-### Key Features
-- **Gymnasium-Compatible:** Standard RL interface for easy integration.
-- **Expanded Dataset:** Trained on **2,100+** synthetic patient scenarios across all 5 MTS categories.
-- **Safety-Aware Rewards:** Heavy penalties for under-triaging critical patients.
-- **Fine-Tuned Agent:** Llama 3.2 3B trained with Unsloth (4-bit QLoRA) - **60% accuracy validated**.
-- **Age-Aware Triage:** Demographic parsing for accurate risk stratification.
-- **A2A Protocol:** Agent-to-Agent evaluation via AgentBeats platform.
-- **Docker Deployment:** Fully containerized for reproducibility.
-- **Dual Mode:** Runs as interactive demo (Gradio) or API server (A2A).
-
-## 🚀 Quick Start
-
-### Run with Docker
-
-```bash
-# Pull the image
-docker pull nursecitizendeveloper/nursesim-triage:latest
-
-# Run in demo mode (Gradio UI)
-docker run -p 7860:7860 nursecitizendeveloper/nursesim-triage:latest
-
-# Run in A2A mode (API only)
-docker run -e MODE=a2a -p 7860:7860 nursecitizendeveloper/nursesim-triage:latest
-```
-
-### Test the A2A Endpoint
-
-```bash
-# Health check
-curl https://nursecitizendeveloper-nursesim-triage-demo.hf.space/health
-
-# Get agent card
-curl https://nursecitizendeveloper-nursesim-triage-demo.hf.space/.well-known/agent-card.json
-
-# Submit a task
-curl -X POST https://nursecitizendeveloper-nursesim-triage-demo.hf.space/process-task \
-  -H "Content-Type: application/json" \
-  -d '{
-    "complaint": "Chest pain",
-    "vitals": {
-      "heart_rate": 110,
-      "blood_pressure": "90/60",
-      "spo2": 94,
-      "temperature": 37.2
-    }
-  }'
-```
-
-## 🏗️ Project Structure
-
-```
-NurseSim-RL/
-├── nursesim_rl/           # Core environment package
-│   ├── __init__.py
-│   ├── TriageEnv.py       # Gymnasium environment
-│   └── PatientGenerator.py # Synthetic patient generation
-├── notebooks/
-│   └── NurseSim_RL_Unsloth_Training.ipynb  # Training notebook
-├── data/
-│   ├── train.jsonl        # Training dataset (500 examples)
-│   └── val.jsonl          # Validation dataset (100 examples)
-├── app.py                 # Gradio demo application
-├── Dockerfile             # For reproducibility
-├── requirements.txt
-└── README.md
-```
-
-## 🚀 Quick Start
-
-### Installation
-
-```bash
-git clone https://github.com/NurseCitizenDeveloper/NurseSim-RL.git
-cd NurseSim-RL
-pip install -r requirements.txt
-```
-
-### Using the Environment
-
-```python
-import gymnasium as gym
-from nursesim_rl import TriageEnv
-
-env = gym.make("NurseSim-Triage-v0")
-obs, info = env.reset()
-
-# Agent takes an action
-action = {"triage_category": 2, "intervention": 1}
-obs, reward, terminated, truncated, info = env.step(action)
-```
-
-### Running the Demo
-
-**Gradio Mode (Human UI):**
-```bash
-export AGENT_MODE=gradio
-export HF_TOKEN=your_hf_token_here
-python app.py
-```
-
-**AgentBeats A2A Mode (Platform Integration):**
-```bash
-export AGENT_MODE=a2a
-export HF_TOKEN=your_hf_token_here
-python agent_main.py
-```
-
-## 🤖 AgentBeats Integration
-
-This agent is fully compatible with the [AgentBeats platform](https://agentbeats.org) for automated agent evaluation via the **Agent-to-Agent (A2A) protocol**.
-
-### Dual-Mode Architecture
-
-The agent supports two deployment modes:
-
-| Mode | Purpose | Entry Point | Port |
-|------|---------|-------------|------|
-| **Gradio** | Human-facing UI for demos | `app.py` | 7860 |
-| **A2A** | Platform integration for automated evaluation | `agent_main.py` | 8080 |
-
-Set the mode via the `AGENT_MODE` environment variable.
-
-### A2A Protocol Compliance
-
-- **Agent Card:** `.well-known/agent-card.json` - Metadata and schemas
-- **Task Processing:** Structured input/output for triage assessments
-- **Lifecycle Methods:** `reset()`, `health_check()`
-- **Protocol Version:** A2A v1.0
-
-### Local Testing with AgentBeats Controller
-
-```bash
-# Install earthshaker SDK
-pip install earthshaker
-
-# Set environment variables
-export HF_TOKEN=your_hf_token_here
-export AGENT_MODE=a2a
-
-# Run the controller
-earthshaker run_ctrl
-
-# Test the agent card endpoint (in another terminal)
-curl http://localhost:8080/.well-known/agent-card.json | jq
-
-# Submit a test task via A2A protocol
-curl -X POST http://localhost:8080/task \
-  -H "Content-Type: application/json" \
-  -d '{
-    "complaint": "Chest pain and shortness of breath",
-    "vitals": {
-      "heart_rate": 120,
-      "blood_pressure": "85/55",
-      "spo2": 89,
-      "temperature": 37.8
-    }
-  }'
-```
-
-### Docker Deployment
-
-**Build:**
-```bash
-docker build -t nursesim-triage:latest .
-```
-
-**Run in A2A Mode:**
-```bash
-docker run -e HF_TOKEN=$HF_TOKEN -e AGENT_MODE=a2a -p 8080:8080 nursesim-triage:latest
-```
-
-**Run in Gradio Mode:**
-```bash
-docker run -e HF_TOKEN=$HF_TOKEN -e AGENT_MODE=gradio -p 7860:7860 nursesim-triage:latest
-```
-
-## 📊 Training Results & Validation
-
-The agent was fine-tuned using **Unsloth** on a Llama 3.2 3B base model with an expanded dataset of ~2,100 clinical scenarios.
-
-### ✅ Performance Metrics (Validated)
-Evaluated on 15 Gold-Standard Clinical Scenarios using GPT-5.2 as a Clinical Judge.
-
-| Metric | Value | Description |
-|--------|-------|-------------|
-| **Accuracy** | **60%** | Exact match with Manchester Triage Categories (1-5) |
-| **Safety** | **70%+** | Pass Rate for critical life-threat detection (Sepsis, Anaphylaxis) |
-| **Training Loss** | 0.19 | Final loss after 300 steps |
-| **Hardware** | NVIDIA A100 | Google Colab |
-| **Training Time** | 25 minutes | Using Unsloth QLoRA |
-
-### 🧠 Key Methodology: Age-Aware Triage
-Our validation revealed that **parsing Age and Gender** from the patient description is critical for accurate risk stratification (e.g., separating "Chest Pain" in a 72M vs 20M). The model effectively learned these demographic risk factors, improving accuracy from 16% to 60%.
-
-See our [W&B Report]([https://api.wandb.ai/links/mrlincs-nursing-citizen-development/yg3hzy60]) for detailed training curves.
-
-## 🩺 Clinical Framework: Manchester Triage System
-
-| Category | Priority | Target Time | Example |
-|----------|----------|-------------|---------|
-| 1 | Immediate | 0 min | Cardiac arrest, Anaphylaxis |
-| 2 | Very Urgent | 10 min | Chest pain, Stroke |
-| 3 | Urgent | 60 min | Abdominal pain, Fractures |
-| 4 | Standard | 120 min | Minor injuries, Mild illness |
-| 5 | Non-Urgent | 240 min | Minor cuts, GP-suitable |
-
-## 📚 Resources
-
-- **Hugging Face Space:** [Try the Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)
-- **Model Card:** [NurseSim-Triage-Llama-3.2-3B](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)
-- **Training Report:** [W&B Dashboard](https://wandb.ai/mrlincs-nursing-citizen-development/huggingface)
-- **Blog Post:** [Training AI Agents for Clinical Triage](https://huggingface.co/blog/NurseCitizenDeveloper/nursesim-rl-training-ai-agents-clinical-triage)
-- **AgentBeats Profile:** [NurseSim-Triage Benchmark](https://agentbeats.dev/ClinyQAi/nursesim-triage)
-- **Leaderboard:** [Community Results](https://github.com/ClinyQAi/NurseSim-Triage-Leaderboard)
-- **Docker Hub:** [nursecitizendeveloper/nursesim-triage](https://hub.docker.com/r/nursecitizendeveloper/nursesim-triage)
-
-## 🤖 AgentBeats Integration
-
-NurseSim-Triage implements the **Agent-to-Agent (A2A) protocol** for automated benchmarking:
-
-### Protocol Details
-- **Version:** a2a/v1.0
-- **Agent Card:** `/.well-known/agent-card.json`
-- **Health Endpoint:** `/health`
-- **Task Endpoint:** `/process-task` (POST)
-
-### Evaluation Metrics
-- **Triage Accuracy** (0-1): Percentage of correct MTS assignments
-- **Safety Score** (0-1): Penalizes dangerous under-triage
-- **Response Quality** (0-1): Clinical reasoning coherence
-- **Response Time** (ms): Computational efficiency
-
-### Submit Your Agent
-1. Register on [AgentBeats](https://agentbeats.dev)
-2. Implement the A2A protocol
-3. Submit to NurseSim-Triage benchmark
-4. View results on the [leaderboard](https://agentbeats.dev/ClinyQAi/nursesim-triage)
-
-## 🐳 Deployment
-
-### Hugging Face Spaces
-Deployed on **NVIDIA T4 (Medium)** GPU with:
-- 4-bit quantization (`BitsAndBytesConfig`)
-- Asynchronous model loading
-- Dual-mode support (Gradio + A2A)
-
-### Docker
-```bash
-# Build locally
-docker build -t nursesim-triage .
-
-# Run in demo mode
-docker run -p 7860:7860 nursesim-triage
-
-# Run in A2A mode
-docker run -e MODE=a2a -p 7860:7860 nursesim-triage
-```
-
-### Environment Variables
-- `MODE`: `gradio` (default) or `a2a`
-- `HF_TOKEN`: Hugging Face API token (for private models)
-- `OMP_NUM_THREADS`: OpenMP threads (auto-configured)
-
-## 🏆 OpenEnv Challenge
-
-This project was submitted to the **OpenEnv Challenge 2026** (Berkeley RDI AgentX-AgentBeats Competition).
-
-**Key Contributions:**
-- Novel benchmark for clinical AI evaluation
-- Safety-focused metrics (penalizes under-triage)
-- Open-source training pipeline
-- Reproducible Docker deployment
-- Community leaderboard
-
-## 📄 License
-
-MIT License - See [LICENSE](LICENSE) for details.
-
-## 🙏 Acknowledgements
-
-**Mentors and Champions of Innovation:**
-- **Dr Clare Cable**, Chief Executive, Burdett Trust for Nursing — For championing Relational Intelligence
-- **Professor Joanne Bosanquet**, Chief Executive, Foundation of Nursing Studies — For championing person-centred nursing
-- **Professor Gemma Stacey**, Programme Director, Nursing Now Challenge — For inspiring global nursing leadership
-- **Aisha Holloway**, Chief Nursing Officer, Scotland — For inspiring excellence
-- **Josie Rudman MBE** — Mutual Mentor & champion of nurse-led innovation
-
-**Research & Education Partners:**
-- **Kumbi Kariwo** — Champion of AI equity and bias mitigation
-- **Rohit Sagoo** — Children's Nurse & Innovator in education and practice
-- **Dr Hellena Habte-Asres** — Big Data Researcher, Nurse & Innovator
-- **Kelly Thobekile Ncube** — Senior Lecturer in Adult Nursing (SFHEA) and Global Health Lecturer Volunteer Fellow
-
-**Technical Community:**
-- **OpenEnv Challenge** — Berkeley RDI, PyTorch, Hugging Face, Unsloth
-- **Manchester Triage System** — Clinical framework
-- **Unsloth AI** — 2x faster fine-tuning
-- **AgentBeats** — A2A protocol infrastructure
-- **NVIDIA** — T4 GPU infrastructure
-
----
-
-**Built for the OpenEnv Challenge 2026** 🏆
+---
+title: NurseSim Triage
+emoji: 🏥
+colorFrom: blue
+colorTo: indigo
+sdk: docker
+pinned: false
+---
+
+# NurseSim-RL: A Healthcare Agent Environment for Clinical Triage
+
+[![AgentBeats A2A](https://img.shields.io/badge/AgentBeats-A2A%20Enabled-purple)](https://agentbeats.dev/ClinyQAi/nursesim-triage)
+
+[![OpenEnv Challenge](https://img.shields.io/badge/OpenEnv-Challenge%202026-blue)](https://rdi.berkeley.edu/agentx-agentbeats)
+[![Hugging Face Model](https://img.shields.io/badge/🤗-Model-yellow)](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)
+[![W&B Report](https://img.shields.io/badge/W%26B-Report-orange)](https://wandb.ai/mrlincs-nursing-citizen-development/huggingface)
+[![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
+
+> **OpenEnv Challenge Entry** | Berkeley RDI AgentX-AgentBeats Competition  
+> A Gymnasium-compatible RL environment for training AI agents to perform clinical triage using the Manchester Triage System (MTS).
+
+![NurseSim Demo](docs/demo.gif)
+
+## 🎯 Overview
+
+**NurseSim-RL** simulates the decision-making process of a Triage Nurse in an Accident & Emergency (A&E) department. The agent must assess patients based on their chief complaint and vital signs, then assign an appropriate triage category (1-5) according to the Manchester Triage System.
+
+### Key Features
+- **Gymnasium-Compatible:** Standard RL interface for easy integration.
+- **Expanded Dataset:** Trained on **2,100+** synthetic patient scenarios across all 5 MTS categories.
+- **Safety-Aware Rewards:** Heavy penalties for under-triaging critical patients.
+- **Fine-Tuned Agent:** Llama 3.2 3B trained with Unsloth (4-bit QLoRA) - **60% accuracy validated**.
+- **NEW: Semantic RL Mode:** NurseEmbed-powered text embeddings for language-conditioned agents.
+- **Age-Aware Triage:** Demographic parsing for accurate risk stratification.
+- **A2A Protocol:** Agent-to-Agent evaluation via AgentBeats platform.
+- **Docker Deployment:** Fully containerized for reproducibility.
+- **Dual Mode:** Runs as interactive demo (Gradio) or API server (A2A).
+
+## 🚀 Quick Start
+
+### Run with Docker
+
+```bash
+# Pull the image
+docker pull nursecitizendeveloper/nursesim-triage:latest
+
+# Run in demo mode (Gradio UI)
+docker run -p 7860:7860 nursecitizendeveloper/nursesim-triage:latest
+
+# Run in A2A mode (API only)
+docker run -e MODE=a2a -p 7860:7860 nursecitizendeveloper/nursesim-triage:latest
+```
+
+### Test the A2A Endpoint
+
+```bash
+# Health check
+curl https://nursecitizendeveloper-nursesim-triage-demo.hf.space/health
+
+# Get agent card
+curl https://nursecitizendeveloper-nursesim-triage-demo.hf.space/.well-known/agent-card.json
+
+# Submit a task
+curl -X POST https://nursecitizendeveloper-nursesim-triage-demo.hf.space/process-task \
+  -H "Content-Type: application/json" \
+  -d '{
+    "complaint": "Chest pain",
+    "vitals": {
+      "heart_rate": 110,
+      "blood_pressure": "90/60",
+      "spo2": 94,
+      "temperature": 37.2
+    }
+  }'
+```
+
+## 🏗️ Project Structure
+
+```
+NurseSim-RL/
+├── nursesim_rl/           # Core environment package
+│   ├── __init__.py
+│   ├── TriageEnv.py       # Gymnasium environment
+│   └── PatientGenerator.py # Synthetic patient generation
+├── notebooks/
+│   └── NurseSim_RL_Unsloth_Training.ipynb  # Training notebook
+├── data/
+│   ├── train.jsonl        # Training dataset (500 examples)
+│   └── val.jsonl          # Validation dataset (100 examples)
+├── app.py                 # Gradio demo application
+├── Dockerfile             # For reproducibility
+├── requirements.txt
+└── README.md
+```
+
+## 🚀 Quick Start
+
+### Installation
+
+```bash
+git clone https://github.com/NurseCitizenDeveloper/NurseSim-RL.git
+cd NurseSim-RL
+pip install -r requirements.txt
+```
+
+### Using the Environment
+
+```python
+import gymnasium as gym
+from nursesim_rl import TriageEnv
+
+env = gym.make("NurseSim-Triage-v0")
+obs, info = env.reset()
+
+# Agent takes an action
+action = {"triage_category": 2, "intervention": 1}
+obs, reward, terminated, truncated, info = env.step(action)
+```
+
+### Running the Demo
+
+**Gradio Mode (Human UI):**
+```bash
+export AGENT_MODE=gradio
+export HF_TOKEN=your_hf_token_here
+python app.py
+```
+
+**AgentBeats A2A Mode (Platform Integration):**
+```bash
+export AGENT_MODE=a2a
+export HF_TOKEN=your_hf_token_here
+python agent_main.py
+```
+
+## 🤖 AgentBeats Integration
+
+This agent is fully compatible with the [AgentBeats platform](https://agentbeats.org) for automated agent evaluation via the **Agent-to-Agent (A2A) protocol**.
+
+### Dual-Mode Architecture
+
+The agent supports two deployment modes:
+
+| Mode | Purpose | Entry Point | Port |
+|------|---------|-------------|------|
+| **Gradio** | Human-facing UI for demos | `app.py` | 7860 |
+| **A2A** | Platform integration for automated evaluation | `agent_main.py` | 8080 |
+
+Set the mode via the `AGENT_MODE` environment variable.
+
+### A2A Protocol Compliance
+
+- **Agent Card:** `.well-known/agent-card.json` - Metadata and schemas
+- **Task Processing:** Structured input/output for triage assessments
+- **Lifecycle Methods:** `reset()`, `health_check()`
+- **Protocol Version:** A2A v1.0
+
+### Local Testing with AgentBeats Controller
+
+```bash
+# Install earthshaker SDK
+pip install earthshaker
+
+# Set environment variables
+export HF_TOKEN=your_hf_token_here
+export AGENT_MODE=a2a
+
+# Run the controller
+earthshaker run_ctrl
+
+# Test the agent card endpoint (in another terminal)
+curl http://localhost:8080/.well-known/agent-card.json | jq
+
+# Submit a test task via A2A protocol
+curl -X POST http://localhost:8080/task \
+  -H "Content-Type: application/json" \
+  -d '{
+    "complaint": "Chest pain and shortness of breath",
+    "vitals": {
+      "heart_rate": 120,
+      "blood_pressure": "85/55",
+      "spo2": 89,
+      "temperature": 37.8
+    }
+  }'
+```
+
+### Docker Deployment
+
+**Build:**
+```bash
+docker build -t nursesim-triage:latest .
+```
+
+**Run in A2A Mode:**
+```bash
+docker run -e HF_TOKEN=$HF_TOKEN -e AGENT_MODE=a2a -p 8080:8080 nursesim-triage:latest
+```
+
+**Run in Gradio Mode:**
+```bash
+docker run -e HF_TOKEN=$HF_TOKEN -e AGENT_MODE=gradio -p 7860:7860 nursesim-triage:latest
+```
+
+## 📊 Training Results & Validation
+
+The agent was fine-tuned using **Unsloth** on a Llama 3.2 3B base model with an expanded dataset of ~2,100 clinical scenarios.
+
+### ✅ Performance Metrics (Validated)
+Evaluated on 15 Gold-Standard Clinical Scenarios using GPT-5.2 as a Clinical Judge.
+
+| Metric | Value | Description |
+|--------|-------|-------------|
+| **Accuracy** | **60%** | Exact match with Manchester Triage Categories (1-5) |
+| **Safety** | **70%+** | Pass Rate for critical life-threat detection (Sepsis, Anaphylaxis) |
+| **Training Loss** | 0.19 | Final loss after 300 steps |
+| **Hardware** | NVIDIA A100 | Google Colab |
+| **Training Time** | 25 minutes | Using Unsloth QLoRA |
+
+### 🧠 Key Methodology: Age-Aware Triage
+Our validation revealed that **parsing Age and Gender** from the patient description is critical for accurate risk stratification (e.g., separating "Chest Pain" in a 72M vs 20M). The model effectively learned these demographic risk factors, improving accuracy from 16% to 60%.
+
+See our [W&B Report](https://wandb.ai/mrlincs-nursing-citizen-development/huggingface) for detailed training curves.
+
+## 🩺 Clinical Framework: Manchester Triage System
+
+| Category | Priority | Target Time | Example |
+|----------|----------|-------------|---------|
+| 1 | Immediate | 0 min | Cardiac arrest, Anaphylaxis |
+| 2 | Very Urgent | 10 min | Chest pain, Stroke |
+| 3 | Urgent | 60 min | Abdominal pain, Fractures |
+| 4 | Standard | 120 min | Minor injuries, Mild illness |
+| 5 | Non-Urgent | 240 min | Minor cuts, GP-suitable |
+
+## 📚 Resources
+
+- **Hugging Face Space:** [Try the Demo](https://huggingface.co/spaces/NurseCitizenDeveloper/NurseSim-Triage-Demo)
+- **Model Card:** [NurseSim-Triage-Llama-3.2-3B](https://huggingface.co/NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B)
+- **Training Report:** [W&B Dashboard](https://wandb.ai/mrlincs-nursing-citizen-development/huggingface)
+- **Blog Post:** [Training AI Agents for Clinical Triage](https://huggingface.co/blog/NurseCitizenDeveloper/nursesim-rl-training-ai-agents-clinical-triage)
+- **AgentBeats Profile:** [NurseSim-Triage Benchmark](https://agentbeats.dev/ClinyQAi/nursesim-triage)
+- **Leaderboard:** [Community Results](https://github.com/ClinyQAi/NurseSim-Triage-Leaderboard)
+- **Docker Hub:** [nursecitizendeveloper/nursesim-triage](https://hub.docker.com/r/nursecitizendeveloper/nursesim-triage)
+
+## 🤖 AgentBeats Integration
+
+NurseSim-Triage implements the **Agent-to-Agent (A2A) protocol** for automated benchmarking:
+
+### Protocol Details
+- **Version:** a2a/v1.0
+- **Agent Card:** `/.well-known/agent-card.json`
+- **Health Endpoint:** `/health`
+- **Task Endpoint:** `/process-task` (POST)
+
+### Evaluation Metrics
+- **Triage Accuracy** (0-1): Percentage of correct MTS assignments
+- **Safety Score** (0-1): Penalizes dangerous under-triage
+- **Response Quality** (0-1): Clinical reasoning coherence
+- **Response Time** (ms): Computational efficiency
+
+### Submit Your Agent
+1. Register on [AgentBeats](https://agentbeats.dev)
+2. Implement the A2A protocol
+3. Submit to NurseSim-Triage benchmark
+4. View results on the [leaderboard](https://agentbeats.dev/ClinyQAi/nursesim-triage)
+
+## 🐳 Deployment
+
+### Hugging Face Spaces
+Deployed on **NVIDIA T4 (Medium)** GPU with:
+- 4-bit quantization (`BitsAndBytesConfig`)
+- Asynchronous model loading
+- Dual-mode support (Gradio + A2A)
+
+### Docker
+```bash
+# Build locally
+docker build -t nursesim-triage .
+
+# Run in demo mode
+docker run -p 7860:7860 nursesim-triage
+
+# Run in A2A mode
+docker run -e MODE=a2a -p 7860:7860 nursesim-triage
+```
+
+### Environment Variables
+- `MODE`: `gradio` (default) or `a2a`
+- `HF_TOKEN`: Hugging Face API token (for private models)
+- `OMP_NUM_THREADS`: OpenMP threads (auto-configured)
+
+## 🏆 OpenEnv Challenge
+
+This project was submitted to the **OpenEnv Challenge 2026** (Berkeley RDI AgentX-AgentBeats Competition).
+
+**Key Contributions:**
+- Novel benchmark for clinical AI evaluation
+- Safety-focused metrics (penalizes under-triage)
+- Open-source training pipeline
+- Reproducible Docker deployment
+- Community leaderboard
+
+## 📄 License
+
+MIT License - See [LICENSE](LICENSE) for details.
+
+## 🙏 Acknowledgements
+
+**Mentors and Champions of Innovation:**
+- **Dr Clare Cable**, Chief Executive, Burdett Trust for Nursing — For championing Relational Intelligence
+- **Professor Joanne Bosanquet**, Chief Executive, Foundation of Nursing Studies — For championing person-centred nursing
+- **Professor Gemma Stacey**, Programme Director, Nursing Now Challenge — For inspiring global nursing leadership
+- **Aisha Holloway**, Chief Nursing Officer, Scotland — For inspiring excellence
+- **Josie Rudman MBE** — Mutual Mentor & champion of nurse-led innovation
+
+**Research & Education Partners:**
+- **Kumbi Kariwo** — Champion of AI equity and bias mitigation
+- **Rohit Sagoo** — Children's Nurse & Innovator in education and practice
+- **Dr Hellena Habte-Asres** — Big Data Researcher, Nurse & Innovator
+- **Kelly Thobekile Ncube** — Senior Lecturer in Adult Nursing (SFHEA) and Global Health Lecturer Volunteer Fellow
+
+**Technical Community:**
+- **OpenEnv Challenge** — Berkeley RDI, PyTorch, Hugging Face, Unsloth
+- **Manchester Triage System** — Clinical framework
+- **Unsloth AI** — 2x faster fine-tuning
+- **AgentBeats** — A2A protocol infrastructure
+- **NVIDIA** — T4 GPU infrastructure
+
+---
+
+**Built for the OpenEnv Challenge 2026** 🏆

SUBMISSION_ABSTRACT.md

@@ -0,0 +1,30 @@
+# Submission Abstract: NurseSim-RL
+
+## Project Name
+NurseSim-RL: A Healthcare Agent Environment for Clinical Triage
+
+## Abstract (for submission form)
+
+NurseSim-RL is a Gymnasium-compatible reinforcement learning environment that simulates clinical triage in an Emergency Department (A&E) setting. The environment challenges AI agents to assess patients based on natural language chief complaints and vital sign data, then assign appropriate triage categories (1-5) according to the Manchester Triage System (MTS).
+
+**Key Contributions:**
+1. **Novel Healthcare RL Environment:** A safety-critical environment where incorrect decisions carry severe penalties, modeling real-world clinical risk.
+2. **Synthetic Clinical Dataset:** 500+ diverse patient scenarios covering all 5 MTS categories, with realistic vital sign variations.
+3. **Fine-Tuned LLM Agent:** A Llama 3.2 3B model trained using Unsloth (4-bit QLoRA) demonstrating rapid domain adaptation (2.8 → 0.08 loss in 100 steps).
+4. **Reproducible Pipeline:** Complete training notebook, Dockerfile, and Gradio demo for immediate deployment.
+
+**Evaluation Focus:** Healthcare Agent Track - The benchmark evaluates clinical reasoning, safety awareness, and resource allocation under time pressure.
+
+**Impact:** This environment enables development and testing of AI agents for healthcare decision support, with direct applications in triage training, clinical education, and NHS workforce optimization.
+
+---
+
+## Suggested Answers for Form Fields
+
+**Participation Category:** Create a new benchmark
+
+**Evaluation Track(s):** Healthcare Agent
+
+**Specific Benchmarks:** N/A (new benchmark)
+
+**Demo Video Title:** "NurseSim-RL: AI Triage Agent Demo - OpenEnv Challenge 2026"

WANDB_REPORT_TEXT.md

@@ -0,0 +1,50 @@
+# NurseSim-RL: Training a Specialist Triage Agent
+**By NurseCitizenDeveloper**
+
+## 🎯 The Mission: OpenEnv Challenge
+The goal of **NurseSim-RL** is to create an AI agent capable of performing safe, accurate clinical triage in a simulated Emergency Department. Using the **Manchester Triage System (MTS)**, the agent must assess patient complaints and vitals to assign priority (Category 1-5).
+
+This report documents the fine-tuning of a **Llama 3.2 3B** model to master this complex clinical reasoning task.
+
+---
+
+## 🏗️ Methodology
+
+### The Model
+We selected **Meta's Llama 3.2 3B Instruct** for its balance of reasoning capability and edge-device efficiency.
+- **Optimization:** We used **Unsloth** for 2x faster training and 60% memory reduction.
+- **Quantization:** 4-bit (QLoRA) to fit within Colab GPU constraints.
+
+### The Dataset
+A synthetic dataset of **500 clinical scenarios** was generated using `PatientGenerator.py`.
+- **Inputs:** Natural language "Chief Complaint" + Vitals (HR, BP, SpO2, Temp).
+- **Outputs:** Triage Category (1-5) + Clinical Rationale.
+
+### Hyperparameters
+- **Rank (r):** 16
+- **Alpha:** 16
+- **Learning Rate:** 2e-4 (Linear Decay)
+- **Batch Size:** 8 (Gradient Accumulation: 4)
+- **Max Steps:** 100
+
+---
+
+## 📈 Training Analysis
+
+### rapid Convergence
+As seen in the training logs, the model demonstrated **exceptional adaptability** to the clinical domain.
+
+*   **Loss Curve:** The training loss plummeted from an initial **2.8** to **<0.1** within just 100 steps (~6 epochs). This indicates that the underlying logic of the Manchester Triage System is highly structured and learnable for a model of this caliber.
+*   **Stability:** The `grad_norm` graph shows initial variance (as the model adjusted to the new format) followed by a smooth stabilization, confirming that the learning rate of 2e-4 was appropriate.
+
+### Why this matters
+The rapid convergence suggests that we successfully turned a general-purpose LLM into a **specialized clinical agent** without needing massive compute. The final low loss score implies the model isn't just guessing—it has internalized the rules of triage.
+
+---
+
+## 🏥 Conclusion & Next Steps
+We have successfully trained a robust Triage Agent.
+- **Status:** The model is now hosted on Hugging Face (`NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B`).
+- **Deployment:** A Gradio web application is being deployed to allow real-time interaction with the agent.
+
+**Verdict:** Llama 3.2 + Unsloth is a viable pipeline for creating lightweight, domain-specific clinical agents.

agent_main.py

@@ -1,272 +1,272 @@
-#!/usr/bin/env python3
-"""
-NurseSim-Triage Hybrid Agent Entry Point
-
-This module combines the A2A API (for AgentBeats) and the Gradio UI (for Human/Demo)
-into a single FastAPI application listening on port 7860.
-"""
-
-import os
-import json
-import torch
-import uvicorn
-import asyncio
-import gradio as gr
-from contextlib import asynccontextmanager
-from fastapi import FastAPI, HTTPException, Request
-from fastapi.responses import JSONResponse
-from fastapi.middleware.cors import CORSMiddleware
-from typing import Dict, Any
-from pydantic import BaseModel
-from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
-from peft import PeftModel
-
-# ==========================================
-# Data Models
-# ==========================================
-
-class Vitals(BaseModel):
-    heart_rate: int = 80
-    blood_pressure: str = "120/80"
-    spo2: int = 98
-    temperature: float = 37.0
-
-class TaskInput(BaseModel):
-    complaint: str
-    vitals: Vitals
-
-# ==========================================
-# Agent Core Logic
-# ==========================================
-
-class NurseSimTriageAgent:
-    """
-    Shared agent logic for both API and UI.
-    """
-    
-    def __init__(self):
-        """Initialize the triage agent placeholder."""
-        self.model = None
-        self.tokenizer = None
-        self.HF_TOKEN = os.environ.get("HF_TOKEN")
-        
-        if not self.HF_TOKEN:
-            print("WARNING: HF_TOKEN not set. Model loading will fail if authentication is required.")
-    
-    async def load_model(self):
-        """Load the base model and LoRA adapters asynchronously."""
-        if self.model is not None:
-            return
-
-        try:
-            print("⏳ Starting model load...")
-            base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
-            adapter_id = "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B"
-            
-            # Offload heavy loading to thread
-            await asyncio.to_thread(self._load_weights, base_model_id, adapter_id)
-            
-            print("✅ Model loaded successfully!")
-        except Exception as e:
-            print(f"❌ CRITICAL ERROR loading model: {e}")
-            import traceback
-            traceback.print_exc()
-
-    def _load_weights(self, base_model_id, adapter_id):
-        print(f"Loading tokenizer from {adapter_id}...")
-        self.tokenizer = AutoTokenizer.from_pretrained(adapter_id, token=self.HF_TOKEN)
-        
-        print(f"Loading base model {base_model_id} with 4-bit quantization...")
-        bnb_config = BitsAndBytesConfig(
-            load_in_4bit=True,
-            bnb_4bit_compute_dtype=torch.float16,
-            bnb_4bit_quant_type="nf4",
-            bnb_4bit_use_double_quant=True,
-        )
-
-        self.model = AutoModelForCausalLM.from_pretrained(
-            base_model_id,
-            quantization_config=bnb_config,
-            device_map="auto",
-            low_cpu_mem_usage=True,
-            token=self.HF_TOKEN,
-        )
-        
-        print(f"Applying LoRA adapters from {adapter_id}...")
-        self.model = PeftModel.from_pretrained(self.model, adapter_id, token=self.HF_TOKEN)
-        self.model.eval()
-
-    def get_response(self, complaint: str, hr: int, bp: str, spo2: int, temp: float) -> str:
-        """Shared inference logic."""
-        if self.model is None:
-            return "⚠️ System is warming up. Please try again in 30 seconds."
-
-        prompt = f"""### Instruction:
-You are an expert A&E Triage Nurse. Assess the following patient and provide your triage decision.
-
-### Input:
-Patient Complaint: {complaint}
-Vitals: HR {hr}, BP {bp}, SpO2 {spo2}%, Temp {temp}C.
-
-### Response:"""
-        
-        inputs = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
-        
-        with torch.no_grad():
-            outputs = self.model.generate(
-                **inputs,
-                max_new_tokens=256,
-                do_sample=True,
-                temperature=0.7,
-                pad_token_id=self.tokenizer.eos_token_id,
-            )
-        
-        response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
-        if "### Response:" in response:
-            response = response.split("### Response:")[-1].strip()
-            
-        return response
-
-    def process_task(self, task: Dict[str, Any]) -> Dict[str, Any]:
-        """Process an API task."""
-        if self.model is None:
-            return {
-                "error": "ModelStillLoading", 
-                "message": "The agent is still warming up. Please retry in 30 seconds."
-            }
-
-        try:
-            complaint = task.get("complaint", "")
-            vitals = task.get("vitals", {})
-            response = self.get_response(
-                complaint, 
-                vitals.get("heart_rate", 80),
-                vitals.get("blood_pressure", "120/80"),
-                vitals.get("spo2", 98),
-                vitals.get("temperature", 37.0)
-            )
-            
-            return {
-                "triage_category": self._extract_triage_category(response),
-                "assessment": response,
-                "recommended_action": self._extract_recommended_action(response)
-            }
-            
-        except Exception as e:
-            return {"error": str(e), "triage_category": "Error"}
-    
-    def _extract_triage_category(self, response: str) -> str:
-        response_lower = response.lower()
-        if "immediate" in response_lower or "resuscitation" in response_lower: return "Immediate"
-        elif "very urgent" in response_lower or "emergency" in response_lower: return "Very Urgent"
-        elif "urgent" in response_lower: return "Urgent"
-        elif "standard" in response_lower: return "Standard"
-        elif "non-urgent" in response_lower or "non urgent" in response_lower: return "Non-Urgent"
-        else: return "Standard"
-    
-    def _extract_recommended_action(self, response: str) -> str:
-        if "monitor" in response.lower(): return "Monitor patient closely"
-        elif "immediate" in response.lower() or "urgent" in response.lower(): return "Immediate medical attention required"
-        else: return "Continue assessment and treatment as per protocol"
-    
-    def health_check(self) -> Dict[str, Any]:
-        return {
-            "status": "healthy" if self.model is not None else "loading",
-            "model_loaded": self.model is not None,
-            "gpu_available": torch.cuda.is_available()
-        }
-
-# ==========================================
-# Application Setup
-# ==========================================
-
-agent = NurseSimTriageAgent()
-
-@asynccontextmanager
-async def lifespan(app: FastAPI):
-    print("🚀 Server starting. Triggering model load task...")
-    asyncio.create_task(agent.load_model())
-    yield
-    print("🛑 Server shutting down.")
-
-app = FastAPI(title="NurseSim-Triage Agent", version="1.2.0", lifespan=lifespan)
-
-app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"],
-    allow_credentials=True,
-    allow_methods=["*"],
-    allow_headers=["*"],
-)
-
-# ==========================================
-# API Endpoints
-# ==========================================
-
-@app.get("/health")
-async def health_check():
-    return agent.health_check()
-
-@app.get("/.well-known/agent-card.json")
-async def get_agent_card():
-    card_path = ".well-known/agent-card.json"
-    if os.path.exists(card_path):
-        with open(card_path, "r") as f:
-            return json.load(f)
-    raise HTTPException(status_code=404, detail="Agent card not found")
-
-@app.post("/process-task")
-async def process_task(task: TaskInput):
-    result = agent.process_task(task.dict())
-    if "error" in result and result.get("message") == "ModelStillLoading":
-        raise HTTPException(status_code=503, detail=result["message"])
-    return result
-
-# ==========================================
-# Gradio UI Integration
-# ==========================================
-
-def gradio_predict(complaint, hr, bp, spo2, temp):
-    return agent.get_response(complaint, hr, bp, spo2, temp)
-
-with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("""
-    # 🩺 NurseSim AI: Emergency Triage Simulator
-    **An AI agent fine-tuned for the Manchester Triage System (MTS).**
-    *Developed for the OpenEnv Challenge by NurseCitizenDeveloper.*
-    
-    > ⚡ **Hybrid Mode**: Serving both Gradio UI and A2A API (AgentBeats)
-    """)
-    
-    with gr.Row():
-        with gr.Column():
-            complaint = gr.Textbox(label="Chief Complaint", placeholder="e.g., Shortness of breath...")
-            with gr.Row():
-                hr = gr.Number(label="Heart Rate", value=80)
-                bp = gr.Textbox(label="Blood Pressure", placeholder="e.g., 120/80")
-            with gr.Row():
-                spo2 = gr.Slider(label="SpO2 (%)", minimum=50, maximum=100, value=98)
-                temp = gr.Number(label="Temperature (C)", value=37.0)
-            
-            submit_btn = gr.Button("Assess Patient", variant="primary")
-            
-        with gr.Column():
-            output_text = gr.Textbox(label="AI Triage Assessment", lines=10)
-            gr.Markdown("### ⚠️ Research Prototype - Not for Clinical Use")
-
-    submit_btn.click(gradio_predict, inputs=[complaint, hr, bp, spo2, temp], outputs=output_text)
-    
-    gr.Examples(
-        examples=[
-            ["Crushing chest pain and nausea", 110, "90/60", 94, 37.2],
-            ["Twisted ankle at football", 75, "125/85", 99, 36.8],
-        ],
-        inputs=[complaint, hr, bp, spo2, temp]
-    )
-
-# Mount Gradio app to FastAPI at root
-app = gr.mount_gradio_app(app, demo, path="/")
-
-if __name__ == "__main__":
-    print("Starting Hybrid Server on port 7860...")
-    uvicorn.run(app, host="0.0.0.0", port=7860)
+#!/usr/bin/env python3
+"""
+NurseSim-Triage Hybrid Agent Entry Point
+
+This module combines the A2A API (for AgentBeats) and the Gradio UI (for Human/Demo)
+into a single FastAPI application listening on port 7860.
+"""
+
+import os
+import json
+import torch
+import uvicorn
+import asyncio
+import gradio as gr
+from contextlib import asynccontextmanager
+from fastapi import FastAPI, HTTPException, Request
+from fastapi.responses import JSONResponse
+from fastapi.middleware.cors import CORSMiddleware
+from typing import Dict, Any
+from pydantic import BaseModel
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+from peft import PeftModel
+
+# ==========================================
+# Data Models
+# ==========================================
+
+class Vitals(BaseModel):
+    heart_rate: int = 80
+    blood_pressure: str = "120/80"
+    spo2: int = 98
+    temperature: float = 37.0
+
+class TaskInput(BaseModel):
+    complaint: str
+    vitals: Vitals
+
+# ==========================================
+# Agent Core Logic
+# ==========================================
+
+class NurseSimTriageAgent:
+    """
+    Shared agent logic for both API and UI.
+    """
+    
+    def __init__(self):
+        """Initialize the triage agent placeholder."""
+        self.model = None
+        self.tokenizer = None
+        self.HF_TOKEN = os.environ.get("HF_TOKEN")
+        
+        if not self.HF_TOKEN:
+            print("WARNING: HF_TOKEN not set. Model loading will fail if authentication is required.")
+    
+    async def load_model(self):
+        """Load the base model and LoRA adapters asynchronously."""
+        if self.model is not None:
+            return
+
+        try:
+            print("⏳ Starting model load...")
+            base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
+            adapter_id = "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B"
+            
+            # Offload heavy loading to thread
+            await asyncio.to_thread(self._load_weights, base_model_id, adapter_id)
+            
+            print("✅ Model loaded successfully!")
+        except Exception as e:
+            print(f"❌ CRITICAL ERROR loading model: {e}")
+            import traceback
+            traceback.print_exc()
+
+    def _load_weights(self, base_model_id, adapter_id):
+        print(f"Loading tokenizer from {adapter_id}...")
+        self.tokenizer = AutoTokenizer.from_pretrained(adapter_id, token=self.HF_TOKEN)
+        
+        print(f"Loading base model {base_model_id} with 4-bit quantization...")
+        bnb_config = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_quant_type="nf4",
+            bnb_4bit_use_double_quant=True,
+        )
+
+        self.model = AutoModelForCausalLM.from_pretrained(
+            base_model_id,
+            quantization_config=bnb_config,
+            device_map="auto",
+            low_cpu_mem_usage=True,
+            token=self.HF_TOKEN,
+        )
+        
+        print(f"Applying LoRA adapters from {adapter_id}...")
+        self.model = PeftModel.from_pretrained(self.model, adapter_id, token=self.HF_TOKEN)
+        self.model.eval()
+
+    def get_response(self, complaint: str, hr: int, bp: str, spo2: int, temp: float) -> str:
+        """Shared inference logic."""
+        if self.model is None:
+            return "⚠️ System is warming up. Please try again in 30 seconds."
+
+        prompt = f"""### Instruction:
+You are an expert A&E Triage Nurse. Assess the following patient and provide your triage decision.
+
+### Input:
+Patient Complaint: {complaint}
+Vitals: HR {hr}, BP {bp}, SpO2 {spo2}%, Temp {temp}C.
+
+### Response:"""
+        
+        inputs = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
+        
+        with torch.no_grad():
+            outputs = self.model.generate(
+                **inputs,
+                max_new_tokens=256,
+                do_sample=True,
+                temperature=0.7,
+                pad_token_id=self.tokenizer.eos_token_id,
+            )
+        
+        response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
+        if "### Response:" in response:
+            response = response.split("### Response:")[-1].strip()
+            
+        return response
+
+    def process_task(self, task: Dict[str, Any]) -> Dict[str, Any]:
+        """Process an API task."""
+        if self.model is None:
+            return {
+                "error": "ModelStillLoading", 
+                "message": "The agent is still warming up. Please retry in 30 seconds."
+            }
+
+        try:
+            complaint = task.get("complaint", "")
+            vitals = task.get("vitals", {})
+            response = self.get_response(
+                complaint, 
+                vitals.get("heart_rate", 80),
+                vitals.get("blood_pressure", "120/80"),
+                vitals.get("spo2", 98),
+                vitals.get("temperature", 37.0)
+            )
+            
+            return {
+                "triage_category": self._extract_triage_category(response),
+                "assessment": response,
+                "recommended_action": self._extract_recommended_action(response)
+            }
+            
+        except Exception as e:
+            return {"error": str(e), "triage_category": "Error"}
+    
+    def _extract_triage_category(self, response: str) -> str:
+        response_lower = response.lower()
+        if "immediate" in response_lower or "resuscitation" in response_lower: return "Immediate"
+        elif "very urgent" in response_lower or "emergency" in response_lower: return "Very Urgent"
+        elif "urgent" in response_lower: return "Urgent"
+        elif "standard" in response_lower: return "Standard"
+        elif "non-urgent" in response_lower or "non urgent" in response_lower: return "Non-Urgent"
+        else: return "Standard"
+    
+    def _extract_recommended_action(self, response: str) -> str:
+        if "monitor" in response.lower(): return "Monitor patient closely"
+        elif "immediate" in response.lower() or "urgent" in response.lower(): return "Immediate medical attention required"
+        else: return "Continue assessment and treatment as per protocol"
+    
+    def health_check(self) -> Dict[str, Any]:
+        return {
+            "status": "healthy" if self.model is not None else "loading",
+            "model_loaded": self.model is not None,
+            "gpu_available": torch.cuda.is_available()
+        }
+
+# ==========================================
+# Application Setup
+# ==========================================
+
+agent = NurseSimTriageAgent()
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    print("🚀 Server starting. Triggering model load task...")
+    asyncio.create_task(agent.load_model())
+    yield
+    print("🛑 Server shutting down.")
+
+app = FastAPI(title="NurseSim-Triage Agent", version="1.2.0", lifespan=lifespan)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# ==========================================
+# API Endpoints
+# ==========================================
+
+@app.get("/health")
+async def health_check():
+    return agent.health_check()
+
+@app.get("/.well-known/agent-card.json")
+async def get_agent_card():
+    card_path = ".well-known/agent-card.json"
+    if os.path.exists(card_path):
+        with open(card_path, "r") as f:
+            return json.load(f)
+    raise HTTPException(status_code=404, detail="Agent card not found")
+
+@app.post("/process-task")
+async def process_task(task: TaskInput):
+    result = agent.process_task(task.dict())
+    if "error" in result and result.get("message") == "ModelStillLoading":
+        raise HTTPException(status_code=503, detail=result["message"])
+    return result
+
+# ==========================================
+# Gradio UI Integration
+# ==========================================
+
+def gradio_predict(complaint, hr, bp, spo2, temp):
+    return agent.get_response(complaint, hr, bp, spo2, temp)
+
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("""
+    # 🩺 NurseSim AI: Emergency Triage Simulator
+    **An AI agent fine-tuned for the Manchester Triage System (MTS).**
+    *Developed for the OpenEnv Challenge by NurseCitizenDeveloper.*
+    
+    > ⚡ **Hybrid Mode**: Serving both Gradio UI and A2A API (AgentBeats)
+    """)
+    
+    with gr.Row():
+        with gr.Column():
+            complaint = gr.Textbox(label="Chief Complaint", placeholder="e.g., Shortness of breath...")
+            with gr.Row():
+                hr = gr.Number(label="Heart Rate", value=80)
+                bp = gr.Textbox(label="Blood Pressure", placeholder="e.g., 120/80")
+            with gr.Row():
+                spo2 = gr.Slider(label="SpO2 (%)", minimum=50, maximum=100, value=98)
+                temp = gr.Number(label="Temperature (C)", value=37.0)
+            
+            submit_btn = gr.Button("Assess Patient", variant="primary")
+            
+        with gr.Column():
+            output_text = gr.Textbox(label="AI Triage Assessment", lines=10)
+            gr.Markdown("### ⚠️ Research Prototype - Not for Clinical Use")
+
+    submit_btn.click(gradio_predict, inputs=[complaint, hr, bp, spo2, temp], outputs=output_text)
+    
+    gr.Examples(
+        examples=[
+            ["Crushing chest pain and nausea", 110, "90/60", 94, 37.2],
+            ["Twisted ankle at football", 75, "125/85", 99, 36.8],
+        ],
+        inputs=[complaint, hr, bp, spo2, temp]
+    )
+
+# Mount Gradio app to FastAPI at root
+app = gr.mount_gradio_app(app, demo, path="/")
+
+if __name__ == "__main__":
+    print("Starting Hybrid Server on port 7860...")
+    uvicorn.run(app, host="0.0.0.0", port=7860)

app.py

@@ -1,159 +1,159 @@
-import gradio as gr
-import spaces
-import torch
-import os
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel
-
-# Get HF token from environment (set as a Space secret)
-HF_TOKEN = os.environ.get("HF_TOKEN")
-
-# Global model/tokenizer
-model = None
-tokenizer = None
-
-def load_model():
-    global model, tokenizer
-    if model is None:
-        base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
-        adapter_id = "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B"
-        
-        tokenizer = AutoTokenizer.from_pretrained(adapter_id, token=HF_TOKEN)
-        
-        # Load base model in 4-bit
-        model = AutoModelForCausalLM.from_pretrained(
-            base_model_id,
-            torch_dtype=torch.float16,
-            device_map="auto",
-            load_in_4bit=True,
-            token=HF_TOKEN,
-        )
-        # Apply LoRA adapters
-        model = PeftModel.from_pretrained(model, adapter_id, token=HF_TOKEN)
-        model.eval()
-    return model, tokenizer
-
-def format_prompt(complaint, hr, bp, spo2, temp, rr, avpu, age, gender, pmh):
-    # Construct History Dictionary (Critical for Model Accuracy)
-    history_dict = {
-        'age': int(age) if age else "Unknown",
-        'gender': gender,
-        'relevant_PMH': pmh if pmh else "None",
-        'time_course': "See complaint"
-    }
-    
-    # Exact Training Data Format
-    input_text = f"""PATIENT PRESENTING TO A&E TRIAGE
-
-Chief Complaint: "{complaint}"
-
-Vitals:
-- HR: {hr} bpm
-- BP: {bp} mmHg
-- SpO2: {spo2}%
-- RR: {rr} /min
-- Temp: {temp}C
-- AVPU: {avpu}
-
-History: {history_dict}
-
-WAITING ROOM: 12 patients | AVAILABLE BEDS: 4
-
-What is your triage decision?"""
-
-    return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
-
-### Instruction:
-You are an expert A&E Triage Nurse using the Manchester Triage System. Assess the following patient and provide your triage decision with clinical reasoning.
-
-### Input:
-{input_text}
-
-### Response:
-"""
-
-@spaces.GPU(duration=120)
-def triage_patient(complaint, age, gender, pmh, hr, bp, spo2, rr, temp, avpu):
-    model, tokenizer = load_model()
-    
-    prompt = format_prompt(complaint, hr, bp, spo2, temp, rr, avpu, age, gender, pmh)
-    
-    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-    
-    with torch.no_grad():
-        outputs = model.generate(
-            **inputs, 
-            max_new_tokens=256,
-            do_sample=True,
-            temperature=0.6,
-            top_p=0.9,
-            pad_token_id=tokenizer.eos_token_id,
-        )
-    
-    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
-    
-    if "### Response:" in response:
-        response = response.split("### Response:")[-1].strip()
-    
-    return response
-
-# Gradio Interface
-with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", neutral_hue="slate")) as demo:
-    gr.Markdown("""
-    # 🏥 NurseSim AI: Emergency Triage Simulator
-    **An AI agent fine-tuned for the Manchester Triage System (MTS).**
-    
-    > **Note:** This model is trained to be **Age-Aware**. A 72-year-old with chest pain is treated differently than a 20-year-old.
-    """)
-    
-    with gr.Row():
-        with gr.Column(scale=1):
-            gr.Markdown("### 1. Patient Demographics")
-            age = gr.Number(label="Age", value=45)
-            gender = gr.Radio(["Male", "Female"], label="Gender", value="Male")
-            pmh = gr.Textbox(label="Medical History (PMH)", placeholder="e.g., Hypertension, Diabetes, Asthma", lines=2)
-            
-            gr.Markdown("### 2. Presentation")
-            complaint = gr.Textbox(label="Chief Complaint", placeholder="e.g., Crushing chest pain radiating to jaw", lines=2)
-            
-        with gr.Column(scale=1):
-            gr.Markdown("### 3. Vital Signs")
-            with gr.Row():
-                hr = gr.Number(label="HR (bpm)", value=80)
-                rr = gr.Number(label="RR (breaths/min)", value=16)
-            with gr.Row():
-                bp = gr.Textbox(label="BP (mmHg)", value="120/80")
-                spo2 = gr.Slider(label="SpO2 (%)", minimum=50, maximum=100, value=98)
-            with gr.Row():
-                temp = gr.Number(label="Temp (C)", value=37.0)
-                avpu = gr.Dropdown(["A", "V", "P", "U"], label="AVPU", value="A")
-            
-            submit_btn = gr.Button("🚨 Assess Patient", variant="primary", size="lg")
-            
-    with gr.Row():
-        output_text = gr.Textbox(label="AI Triage Assessment", lines=8, show_copy_button=True)
-        
-    gr.Markdown("""
-    ### ⚠️ Safety Disclaimer
-    This system is a **research prototype** developed for the OpenEnv Challenge. 
-    It is **NOT** a certified medical device and should not be used for real clinical decision-making.
-    """)
-
-    submit_btn.click(
-        fn=triage_patient,
-        inputs=[complaint, age, gender, pmh, hr, bp, spo2, rr, temp, avpu],
-        outputs=output_text
-    )
-    
-    gr.Examples(
-        examples=[
-            ["Crushing chest pain and nausea", 72, "Male", "HTN, High Cholesterol", 110, "90/60", 94, 24, 37.2, "A"],
-            ["Twisted ankle at football", 22, "Male", "None", 75, "125/85", 99, 14, 36.8, "A"],
-            ["Swollen tongue after peanuts", 25, "Female", "Nut Allergy", 120, "90/60", 91, 28, 37.5, "A"],
-        ],
-        inputs=[complaint, age, gender, pmh, hr, bp, spo2, rr, temp, avpu],
-        label="Test Scenarios"
-    )
-
-if __name__ == "__main__":
-    demo.launch()
+import gradio as gr
+import spaces
+import torch
+import os
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+
+# Get HF token from environment (set as a Space secret)
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+# Global model/tokenizer
+model = None
+tokenizer = None
+
+def load_model():
+    global model, tokenizer
+    if model is None:
+        base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
+        adapter_id = "NurseCitizenDeveloper/NurseSim-Triage-Llama-3.2-3B"
+        
+        tokenizer = AutoTokenizer.from_pretrained(adapter_id, token=HF_TOKEN)
+        
+        # Load base model in 4-bit
+        model = AutoModelForCausalLM.from_pretrained(
+            base_model_id,
+            torch_dtype=torch.float16,
+            device_map="auto",
+            load_in_4bit=True,
+            token=HF_TOKEN,
+        )
+        # Apply LoRA adapters
+        model = PeftModel.from_pretrained(model, adapter_id, token=HF_TOKEN)
+        model.eval()
+    return model, tokenizer
+
+def format_prompt(complaint, hr, bp, spo2, temp, rr, avpu, age, gender, pmh):
+    # Construct History Dictionary (Critical for Model Accuracy)
+    history_dict = {
+        'age': int(age) if age else "Unknown",
+        'gender': gender,
+        'relevant_PMH': pmh if pmh else "None",
+        'time_course': "See complaint"
+    }
+    
+    # Exact Training Data Format
+    input_text = f"""PATIENT PRESENTING TO A&E TRIAGE
+
+Chief Complaint: "{complaint}"
+
+Vitals:
+- HR: {hr} bpm
+- BP: {bp} mmHg
+- SpO2: {spo2}%
+- RR: {rr} /min
+- Temp: {temp}C
+- AVPU: {avpu}
+
+History: {history_dict}
+
+WAITING ROOM: 12 patients | AVAILABLE BEDS: 4
+
+What is your triage decision?"""
+
+    return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
+
+### Instruction:
+You are an expert A&E Triage Nurse using the Manchester Triage System. Assess the following patient and provide your triage decision with clinical reasoning.
+
+### Input:
+{input_text}
+
+### Response:
+"""
+
+@spaces.GPU(duration=120)
+def triage_patient(complaint, age, gender, pmh, hr, bp, spo2, rr, temp, avpu):
+    model, tokenizer = load_model()
+    
+    prompt = format_prompt(complaint, hr, bp, spo2, temp, rr, avpu, age, gender, pmh)
+    
+    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs, 
+            max_new_tokens=256,
+            do_sample=True,
+            temperature=0.6,
+            top_p=0.9,
+            pad_token_id=tokenizer.eos_token_id,
+        )
+    
+    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    
+    if "### Response:" in response:
+        response = response.split("### Response:")[-1].strip()
+    
+    return response
+
+# Gradio Interface
+with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", neutral_hue="slate")) as demo:
+    gr.Markdown("""
+    # 🏥 NurseSim AI: Emergency Triage Simulator
+    **An AI agent fine-tuned for the Manchester Triage System (MTS).**
+    
+    > **Note:** This model is trained to be **Age-Aware**. A 72-year-old with chest pain is treated differently than a 20-year-old.
+    """)
+    
+    with gr.Row():
+        with gr.Column(scale=1):
+            gr.Markdown("### 1. Patient Demographics")
+            age = gr.Number(label="Age", value=45)
+            gender = gr.Radio(["Male", "Female"], label="Gender", value="Male")
+            pmh = gr.Textbox(label="Medical History (PMH)", placeholder="e.g., Hypertension, Diabetes, Asthma", lines=2)
+            
+            gr.Markdown("### 2. Presentation")
+            complaint = gr.Textbox(label="Chief Complaint", placeholder="e.g., Crushing chest pain radiating to jaw", lines=2)
+            
+        with gr.Column(scale=1):
+            gr.Markdown("### 3. Vital Signs")
+            with gr.Row():
+                hr = gr.Number(label="HR (bpm)", value=80)
+                rr = gr.Number(label="RR (breaths/min)", value=16)
+            with gr.Row():
+                bp = gr.Textbox(label="BP (mmHg)", value="120/80")
+                spo2 = gr.Slider(label="SpO2 (%)", minimum=50, maximum=100, value=98)
+            with gr.Row():
+                temp = gr.Number(label="Temp (C)", value=37.0)
+                avpu = gr.Dropdown(["A", "V", "P", "U"], label="AVPU", value="A")
+            
+            submit_btn = gr.Button("🚨 Assess Patient", variant="primary", size="lg")
+            
+    with gr.Row():
+        output_text = gr.Textbox(label="AI Triage Assessment", lines=8, show_copy_button=True)
+        
+    gr.Markdown("""
+    ### ⚠️ Safety Disclaimer
+    This system is a **research prototype** developed for the OpenEnv Challenge. 
+    It is **NOT** a certified medical device and should not be used for real clinical decision-making.
+    """)
+
+    submit_btn.click(
+        fn=triage_patient,
+        inputs=[complaint, age, gender, pmh, hr, bp, spo2, rr, temp, avpu],
+        outputs=output_text
+    )
+    
+    gr.Examples(
+        examples=[
+            ["Crushing chest pain and nausea", 72, "Male", "HTN, High Cholesterol", 110, "90/60", 94, 24, 37.2, "A"],
+            ["Twisted ankle at football", 22, "Male", "None", 75, "125/85", 99, 14, 36.8, "A"],
+            ["Swollen tongue after peanuts", 25, "Female", "Nut Allergy", 120, "90/60", 91, 28, 37.5, "A"],
+        ],
+        inputs=[complaint, age, gender, pmh, hr, bp, spo2, rr, temp, avpu],
+        label="Test Scenarios"
+    )
+
+if __name__ == "__main__":
+    demo.launch()

data/gpt_scenarios.json

@@ -0,0 +1,2892 @@
+{
+  "1": [
+    {
+      "chief_complaint": "I can't breathe and my throat feels tight. I think it was the nuts I ate.",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 85,
+        "rr": 30,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 22,
+        "gender": "male",
+        "relevant_PMH": "peanut allergy",
+        "time_course": "10 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He's not responding. He just fell over.",
+      "vitals": {
+        "hr": 122,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 92,
+        "rr": 28,
+        "temp": 36.8,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 58,
+        "gender": "male",
+        "relevant_PMH": "hypertension, smoker",
+        "time_course": "5 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My baby's skin is turning purple. She's burning up and not waking up!",
+      "vitals": {
+        "hr": 180,
+        "bp_sys": 60,
+        "bp_dia": 30,
+        "spo2": 88,
+        "rr": 40,
+        "temp": 40.3,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 1,
+        "gender": "female",
+        "relevant_PMH": "None",
+        "time_course": "1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "I have chest pain like something is crushing me.",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 100,
+        "bp_dia": 60,
+        "spo2": 85,
+        "rr": 25,
+        "temp": 37.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 68,
+        "gender": "female",
+        "relevant_PMH": "previous heart attack",
+        "time_course": "20 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "I can't stop coughing and I'm struggling to breathe.",
+      "vitals": {
+        "hr": 138,
+        "bp_sys": 90,
+        "bp_dia": 55,
+        "spo2": 82,
+        "rr": 32,
+        "temp": 38.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "non-binary",
+        "relevant_PMH": "asthma",
+        "time_course": "15 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He was lifting something heavy and suddenly started gasping for air.",
+      "vitals": {
+        "hr": 150,
+        "bp_sys": 60,
+        "bp_dia": 35,
+        "spo2": 78,
+        "rr": 40,
+        "temp": 36.7,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 37,
+        "gender": "male",
+        "relevant_PMH": "no known conditions",
+        "time_course": "10 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My pregnant wife's arms and legs are shaking uncontrollably!",
+      "vitals": {
+        "hr": 98,
+        "bp_sys": 160,
+        "bp_dia": 110,
+        "spo2": 93,
+        "rr": 24,
+        "temp": 37.0,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 29,
+        "gender": "female",
+        "relevant_PMH": "34 weeks pregnant",
+        "time_course": "5 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He just collapsed and isn't breathing!",
+      "vitals": {
+        "hr": 0,
+        "bp_sys": 0,
+        "bp_dia": 0,
+        "spo2": 0,
+        "rr": 0,
+        "temp": 36.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 75,
+        "gender": "male",
+        "relevant_PMH": "history of heart failure",
+        "time_course": "3 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He keeps having these fits and won't stop!",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 100,
+        "bp_dia": 60,
+        "spo2": 95,
+        "rr": 20,
+        "temp": 37.2,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 17,
+        "gender": "male",
+        "relevant_PMH": "epilepsy",
+        "time_course": "25 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "She's been vomiting blood and now she's passed out!",
+      "vitals": {
+        "hr": 145,
+        "bp_sys": 65,
+        "bp_dia": 40,
+        "spo2": 90,
+        "rr": 22,
+        "temp": 36.0,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 68,
+        "gender": "female",
+        "relevant_PMH": "peptic ulcer disease",
+        "time_course": "2 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "He's not breathing! He just collapsed, and he's turning blue!",
+      "vitals": {
+        "hr": 30,
+        "bp_sys": 50,
+        "bp_dia": 30,
+        "spo2": 70,
+        "rr": 0,
+        "temp": 35.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 72,
+        "gender": "male",
+        "relevant_PMH": "history of cardiac issues",
+        "time_course": "Collapsed suddenly 5 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "I can't breathe and my throat's closing up after eating peanuts!",
+      "vitals": {
+        "hr": 145,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 82,
+        "rr": 30,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 25,
+        "gender": "female",
+        "relevant_PMH": "known nut allergy",
+        "time_course": "Ate peanuts 10 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "My wife's having a seizure and it's not stopping!",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 135,
+        "bp_dia": 85,
+        "spo2": 95,
+        "rr": 24,
+        "temp": 37.5,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 30,
+        "gender": "female",
+        "relevant_PMH": "epilepsy",
+        "time_course": "Seizure started 25 minutes ago and hasn't stopped."
+      }
+    },
+    {
+      "chief_complaint": "He's not waking up, he just slumped over while eating!",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 92,
+        "rr": 32,
+        "temp": 37.9,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 56,
+        "gender": "male",
+        "relevant_PMH": "type 2 diabetes",
+        "time_course": "Collapsed 15 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "She won't wake up and she's burning up!",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 90,
+        "bp_dia": 55,
+        "spo2": 89,
+        "rr": 28,
+        "temp": 39.8,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 5,
+        "gender": "female",
+        "relevant_PMH": "recent flu-like symptoms",
+        "time_course": "Unresponsive for 20 minutes."
+      }
+    },
+    {
+      "chief_complaint": "I'm choking and can't catch my breath!",
+      "vitals": {
+        "hr": 160,
+        "bp_sys": 85,
+        "bp_dia": 50,
+        "spo2": 80,
+        "rr": 35,
+        "temp": 38.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMH": "smoker",
+        "time_course": "Onset of symptoms 15 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "Mom's bleeding a lot, and she's really dizzy!",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 88,
+        "rr": 28,
+        "temp": 37.2,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 62,
+        "gender": "female",
+        "relevant_PMH": "peptic ulcer",
+        "time_course": "Started vomiting blood 30 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "I found him floating face down in the pool!",
+      "vitals": {
+        "hr": 56,
+        "bp_sys": 100,
+        "bp_dia": 70,
+        "spo2": 88,
+        "rr": 6,
+        "temp": 34.0,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 8,
+        "gender": "male",
+        "relevant_PMH": "previous drowning incident",
+        "time_course": "Discovered in pool 5 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "My husband can't move his right side or speak suddenly!",
+      "vitals": {
+        "hr": 105,
+        "bp_sys": 175,
+        "bp_dia": 95,
+        "spo2": 94,
+        "rr": 18,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 67,
+        "gender": "male",
+        "relevant_PMH": "hypertension",
+        "time_course": "Started 20 minutes ago."
+      }
+    },
+    {
+      "chief_complaint": "She's been having these fits for a while now and is not getting better!",
+      "vitals": {
+        "hr": 160,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 90,
+        "rr": 30,
+        "temp": 38.4,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 19,
+        "gender": "female",
+        "relevant_PMH": "none known",
+        "time_course": "Seizures ongoing for 30 minutes."
+      }
+    },
+    {
+      "chief_complaint": "He's not breathing and turning blue!",
+      "vitals": {
+        "hr": 30,
+        "bp_sys": 60,
+        "bp_dia": 40,
+        "spo2": 50,
+        "rr": 4,
+        "temp": 36.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 72,
+        "gender": "male",
+        "relevant_PMH": "heart failure",
+        "time_course": "collapsed suddenly 5 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "Her skin is covered in a rash and she's struggling to breathe!",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 82,
+        "rr": 38,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "female",
+        "relevant_PMH": "peanut allergy",
+        "time_course": "ate a meal containing peanuts 15 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He was shaking uncontrollably and now he's not waking up!",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 90,
+        "rr": 22,
+        "temp": 37.0,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 8,
+        "gender": "male",
+        "relevant_PMH": "epilepsy",
+        "time_course": "had a seizure 10 minutes ago, still unresponsive"
+      }
+    },
+    {
+      "chief_complaint": "He got hit by a car and there's blood everywhere!",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 88,
+        "rr": 30,
+        "temp": 35.8,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "struck by vehicle 20 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "She's been vomiting blood and now she's not responding!",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 60,
+        "bp_dia": 40,
+        "spo2": 92,
+        "rr": 24,
+        "temp": 37.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 67,
+        "gender": "female",
+        "relevant_PMH": "alcoholic liver disease",
+        "time_course": "vomiting for 2 hours before collapse"
+      }
+    },
+    {
+      "chief_complaint": "I found him floating in the pool and he's not breathing!",
+      "vitals": {
+        "hr": 20,
+        "bp_sys": 50,
+        "bp_dia": 30,
+        "spo2": 60,
+        "rr": 0,
+        "temp": 35.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 16,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "unresponsive for unknown duration"
+      }
+    },
+    {
+      "chief_complaint": "He can't breathe properly and his chest is all tight!",
+      "vitals": {
+        "hr": 150,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 85,
+        "rr": 40,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 35,
+        "gender": "male",
+        "relevant_PMH": "asthma",
+        "time_course": "started suddenly 15 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "She suddenly has a severe headache and can't move her right side!",
+      "vitals": {
+        "hr": 98,
+        "bp_sys": 185,
+        "bp_dia": 110,
+        "spo2": 95,
+        "rr": 22,
+        "temp": 36.9,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 55,
+        "gender": "female",
+        "relevant_PMH": "hypertension",
+        "time_course": "started 20 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "She's been fitting continuously for over five minutes!",
+      "vitals": {
+        "hr": 145,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 89,
+        "rr": 28,
+        "temp": 37.3,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 32,
+        "gender": "female",
+        "relevant_PMH": "bipolar disorder, takes lithium",
+        "time_course": "started seizing 7 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He was just burning some rubbish and he inhaled a lot of smoke!",
+      "vitals": {
+        "hr": 136,
+        "bp_sys": 85,
+        "bp_dia": 55,
+        "spo2": 78,
+        "rr": 32,
+        "temp": 36.4,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 40,
+        "gender": "male",
+        "relevant_PMH": "smoker",
+        "time_course": "exposure happened 10 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My son's not waking up and his skin has purple spots.",
+      "vitals": {
+        "hr": 150,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 88,
+        "rr": 35,
+        "temp": 39.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 6,
+        "gender": "male",
+        "relevant_pmh": "none",
+        "time_course": "Fever for 2 days, not waking up in last hour"
+      }
+    },
+    {
+      "chief_complaint": "She was coughing, now she can't breathe after eating peanuts!",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 75,
+        "bp_dia": 45,
+        "spo2": 85,
+        "rr": 40,
+        "temp": 36.8,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 21,
+        "gender": "female",
+        "relevant_pmh": "peanut allergy",
+        "time_course": "Allergic reaction 10 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My wife is having a seizure and she's pregnant!",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 180,
+        "bp_dia": 110,
+        "spo2": 92,
+        "rr": 22,
+        "temp": 37.2,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 30,
+        "gender": "female",
+        "relevant_pmh": "pregnancy 32 weeks",
+        "time_course": "Seizure started 5 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He just collapsed and isn't breathing!",
+      "vitals": {
+        "hr": 0,
+        "bp_sys": 0,
+        "bp_dia": 0,
+        "spo2": 0,
+        "rr": 0,
+        "temp": 36.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 55,
+        "gender": "male",
+        "relevant_pmh": "previous heart attack",
+        "time_course": "Collapsed 2 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He cut himself badly and is losing a lot of blood!",
+      "vitals": {
+        "hr": 125,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 95,
+        "rr": 28,
+        "temp": 35.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 40,
+        "gender": "male",
+        "relevant_pmh": "none",
+        "time_course": "Accident 10 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He's been vomiting and now he's not waking up!",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 94,
+        "rr": 30,
+        "temp": 37.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 25,
+        "gender": "male",
+        "relevant_pmh": "Type 1 diabetes",
+        "time_course": "Vomiting began 8 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "He's choking and can't breathe after the car crash!",
+      "vitals": {
+        "hr": 150,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 82,
+        "rr": 45,
+        "temp": 36.5,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 35,
+        "gender": "male",
+        "relevant_pmh": "none",
+        "time_course": "Accident 5 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "He's been shivering and now he won't wake up!",
+      "vitals": {
+        "hr": 145,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 90,
+        "rr": 20,
+        "temp": 40.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 72,
+        "gender": "male",
+        "relevant_pmh": "chronic kidney disease",
+        "time_course": "Fever for a day, unresponsive now"
+      }
+    },
+    {
+      "chief_complaint": "He's struggling to breathe, and his lips are turning blue!",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 95,
+        "bp_dia": 55,
+        "spo2": 78,
+        "rr": 34,
+        "temp": 37.8,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 60,
+        "gender": "male",
+        "relevant_pmh": "COPD",
+        "time_course": "Breathing difficulty started 1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "She passed out after having a seizure in the water!",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 85,
+        "bp_dia": 50,
+        "spo2": 89,
+        "rr": 26,
+        "temp": 33.0,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 45,
+        "gender": "female",
+        "relevant_pmh": "epilepsy",
+        "time_course": "Seizure 10 minutes ago, fell into pool"
+      }
+    },
+    {
+      "chief_complaint": "He's not breathing and is turning blue!",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 85,
+        "bp_dia": 50,
+        "spo2": 85,
+        "rr": 8,
+        "temp": 36.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 60,
+        "gender": "male",
+        "relevant_PMH": "history of COPD and smoking",
+        "time_course": "Found unconscious at home with sudden respiratory distress."
+      }
+    },
+    {
+      "chief_complaint": "He's collapsed and he's not waking up!",
+      "vitals": {
+        "hr": 45,
+        "bp_sys": 60,
+        "bp_dia": 40,
+        "spo2": 70,
+        "rr": 4,
+        "temp": 35.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 28,
+        "gender": "male",
+        "relevant_PMH": "history of depression, possible overdose",
+        "time_course": "Found unresponsive by his partner after taking pills."
+      }
+    },
+    {
+      "chief_complaint": "She keeps having seizures and isn't waking up!",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 90,
+        "bp_dia": 50,
+        "spo2": 92,
+        "rr": 20,
+        "temp": 37.8,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 35,
+        "gender": "female",
+        "relevant_PMH": "epilepsy",
+        "time_course": "Has had 5 continuous seizures lasting over 30 minutes."
+      }
+    },
+    {
+      "chief_complaint": "He's bleeding everywhere and can't stand up!",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 94,
+        "rr": 22,
+        "temp": 36.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 50,
+        "gender": "male",
+        "relevant_PMH": "alcohol use disorder",
+        "time_course": "Sudden onset of massive vomiting of blood at home."
+      }
+    },
+    {
+      "chief_complaint": "He's not breathing. He was underwater for a long time!",
+      "vitals": {
+        "hr": 50,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 80,
+        "rr": 0,
+        "temp": 34.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 14,
+        "gender": "male",
+        "relevant_PMH": "asthma",
+        "time_course": "Pulled from a swimming pool after several minutes underwater."
+      }
+    },
+    {
+      "chief_complaint": "Her whole body is swollen, and she can't breathe!",
+      "vitals": {
+        "hr": 150,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 85,
+        "rr": 30,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 22,
+        "gender": "female",
+        "relevant_PMH": "known peanut allergy",
+        "time_course": "Ate dessert containing peanuts, severe reaction in minutes."
+      }
+    },
+    {
+      "chief_complaint": "He's unconscious and his skin looks blotchy!",
+      "vitals": {
+        "hr": 145,
+        "bp_sys": 75,
+        "bp_dia": 40,
+        "spo2": 88,
+        "rr": 28,
+        "temp": 40.0,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 3,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "Suddenly developed a fever, rash, and became unresponsive."
+      }
+    },
+    {
+      "chief_complaint": "She's unconscious and breathing fast!",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 92,
+        "rr": 34,
+        "temp": 37.2,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 24,
+        "gender": "female",
+        "relevant_PMH": "type 1 diabetes",
+        "time_course": "Found with high blood sugar levels after missing insulin doses."
+      }
+    },
+    {
+      "chief_complaint": "She's fainted and keeps shaking!",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 170,
+        "bp_dia": 110,
+        "spo2": 95,
+        "rr": 25,
+        "temp": 37.5,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 30,
+        "gender": "female",
+        "relevant_PMH": "32 weeks pregnant, history of high blood pressure",
+        "time_course": "Sudden onset of seizures with high blood pressure."
+      }
+    },
+    {
+      "chief_complaint": "He's not moving, hit really hard!",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 85,
+        "bp_dia": 55,
+        "spo2": 90,
+        "rr": 28,
+        "temp": 36.5,
+        "avpu": "U"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "Involved in a high-speed car accident, unresponsive at the scene."
+      }
+    }
+  ],
+  "2": [
+    {
+      "chief_complaint": "My chest really hurts, like something's squeezing it.",
+      "vitals": {
+        "hr": 122,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 94,
+        "rr": 20,
+        "temp": 37.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 64,
+        "gender": "male",
+        "relevant_PMH": "hypertension, smoker",
+        "time_course": "30 minutes"
+      }
+    },
+    {
+      "chief_complaint": "I suddenly have the worst headache of my life.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 180,
+        "bp_dia": 110,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 55,
+        "gender": "female",
+        "relevant_PMH": "migraines",
+        "time_course": "15 minutes"
+      }
+    },
+    {
+      "chief_complaint": "I feel really confused and my left side feels weird and weak.",
+      "vitals": {
+        "hr": 76,
+        "bp_sys": 150,
+        "bp_dia": 90,
+        "spo2": 96,
+        "rr": 16,
+        "temp": 36.7,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 72,
+        "gender": "male",
+        "relevant_PMH": "atrial fibrillation",
+        "time_course": "1 hour"
+      }
+    },
+    {
+      "chief_complaint": "My toddler is burning up and won't respond properly.",
+      "vitals": {
+        "hr": 160,
+        "bp_sys": 70,
+        "bp_dia": 40,
+        "spo2": 92,
+        "rr": 30,
+        "temp": 40.2,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 2,
+        "gender": "female",
+        "relevant_PMH": "none",
+        "time_course": "3 hours"
+      }
+    },
+    {
+      "chief_complaint": "I fell off a ladder and now my back really hurts.",
+      "vitals": {
+        "hr": 105,
+        "bp_sys": 100,
+        "bp_dia": 65,
+        "spo2": 95,
+        "rr": 22,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 47,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "30 minutes"
+      }
+    },
+    {
+      "chief_complaint": "I'm in agony, my stomach is killing me!",
+      "vitals": {
+        "hr": 115,
+        "bp_sys": 85,
+        "bp_dia": 55,
+        "spo2": 93,
+        "rr": 24,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 58,
+        "gender": "female",
+        "relevant_PMH": "gallstones",
+        "time_course": "2 hours"
+      }
+    },
+    {
+      "chief_complaint": "I can't feel my legs after diving into the shallow end.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 140,
+        "bp_dia": 85,
+        "spo2": 98,
+        "rr": 18,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 23,
+        "gender": "non-binary",
+        "relevant_PMH": "none",
+        "time_course": "10 minutes"
+      }
+    },
+    {
+      "chief_complaint": "He was shaking all over, but it just stopped.",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 98,
+        "bp_dia": 65,
+        "spo2": 90,
+        "rr": 24,
+        "temp": 38.0,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 9,
+        "gender": "male",
+        "relevant_PMH": "epilepsy",
+        "time_course": "5 minutes"
+      }
+    },
+    {
+      "chief_complaint": "I'm so dizzy, I can't stand up without falling.",
+      "vitals": {
+        "hr": 102,
+        "bp_sys": 95,
+        "bp_dia": 60,
+        "spo2": 93,
+        "rr": 18,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 42,
+        "gender": "female",
+        "relevant_PMH": "diabetes",
+        "time_course": "1 hour"
+      }
+    },
+    {
+      "chief_complaint": "I have this sharp pain in my chest and I feel breathless.",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 85,
+        "bp_dia": 55,
+        "spo2": 91,
+        "rr": 28,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 36,
+        "gender": "transgender female",
+        "relevant_PMH": "asthma",
+        "time_course": "2 hours"
+      }
+    },
+    {
+      "chief_complaint": "My chest is killing me. It started suddenly when I was watching TV.",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 90,
+        "bp_dia": 60,
+        "spo2": 94,
+        "rr": 28,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 67,
+        "gender": "male",
+        "relevant_PMhistory": "hypertension, smoker",
+        "time_course": "30 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "I feel like I've been hit with the worst headache ever, like a thunderclap in my head.",
+      "vitals": {
+        "hr": 100,
+        "bp_sys": 160,
+        "bp_dia": 100,
+        "spo2": 98,
+        "rr": 22,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 54,
+        "gender": "female",
+        "relevant_PMhistory": "migraine sufferer",
+        "time_course": "15 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "I suddenly can't move my right arm and my face feels funny.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 170,
+        "bp_dia": 110,
+        "spo2": 95,
+        "rr": 18,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 59,
+        "gender": "female",
+        "relevant_PMhistory": "type 2 diabetes",
+        "time_course": "1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "My baby is burning up and really drowsy. It's scaring me.",
+      "vitals": {
+        "hr": 180,
+        "bp_sys": 80,
+        "bp_dia": 50,
+        "spo2": 92,
+        "rr": 40,
+        "temp": 40.2,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 2,
+        "gender": "male",
+        "relevant_PMhistory": "none",
+        "time_course": "fever for 2 days, worsened in the last hour"
+      }
+    },
+    {
+      "chief_complaint": "I fell down the stairs and can't feel my legs.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 108,
+        "bp_dia": 72,
+        "spo2": 97,
+        "rr": 20,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 35,
+        "gender": "male",
+        "relevant_PMhistory": "none",
+        "time_course": "30 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "I'm in agony from my stomach. It feels like something's gonna burst.",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 85,
+        "bp_dia": 55,
+        "spo2": 96,
+        "rr": 24,
+        "temp": 38.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 40,
+        "gender": "female",
+        "relevant_PMhistory": "prior appendectomy",
+        "time_course": "started 2 hours ago, worsening"
+      }
+    },
+    {
+      "chief_complaint": "I had a seizure and now I'm confused and weak on one side.",
+      "vitals": {
+        "hr": 102,
+        "bp_sys": 140,
+        "bp_dia": 85,
+        "spo2": 98,
+        "rr": 22,
+        "temp": 37.5,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 28,
+        "gender": "female",
+        "relevant_PMhistory": "epilepsy",
+        "time_course": "seizure 20 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My son was hit by a car. He's awake but not responding well.",
+      "vitals": {
+        "hr": 112,
+        "bp_sys": 105,
+        "bp_dia": 70,
+        "spo2": 93,
+        "rr": 25,
+        "temp": 36.6,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 12,
+        "gender": "male",
+        "relevant_PMhistory": "none",
+        "time_course": "accident 15 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "I've got a splitting headache and can't stop vomiting.",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 150,
+        "bp_dia": 95,
+        "spo2": 97,
+        "rr": 20,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMhistory": "high blood pressure",
+        "time_course": "started suddenly 1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "My whole back feels like it's on fire after lifting a heavy box.",
+      "vitals": {
+        "hr": 100,
+        "bp_sys": 132,
+        "bp_dia": 85,
+        "spo2": 96,
+        "rr": 18,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 50,
+        "gender": "female",
+        "relevant_PMhistory": "chronic lower back pain",
+        "time_course": "1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "It feels like an elephant is sitting on my chest!",
+      "vitals": {
+        "hr": 112,
+        "bp_sys": 92,
+        "bp_dia": 60,
+        "spo2": 89,
+        "rr": 28,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 54,
+        "gender": "male",
+        "relevant_PMH": "hypertension, smoker",
+        "time_course": "30 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My head... it's like a bolt of lightning hit me!",
+      "vitals": {
+        "hr": 84,
+        "bp_sys": 145,
+        "bp_dia": 88,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "female",
+        "relevant_PMH": "migraine",
+        "time_course": "10 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "I can't feel my right arm and leg.",
+      "vitals": {
+        "hr": 78,
+        "bp_sys": 150,
+        "bp_dia": 95,
+        "spo2": 96,
+        "rr": 20,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 67,
+        "gender": "male",
+        "relevant_PMH": "diabetes, previous TIA",
+        "time_course": "started 1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "My son is so hot and won't stop crying.",
+      "vitals": {
+        "hr": 140,
+        "bp_sys": 100,
+        "bp_dia": 65,
+        "spo2": 94,
+        "rr": 40,
+        "temp": 40.2,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 2,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "since last night"
+      }
+    },
+    {
+      "chief_complaint": "I fell off my bike and my back is killing me!",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 105,
+        "bp_dia": 70,
+        "spo2": 98,
+        "rr": 22,
+        "temp": 37.3,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 21,
+        "gender": "female",
+        "relevant_PMH": "none",
+        "time_course": "30 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My stomach is in agony, I can't even move!",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 95,
+        "bp_dia": 60,
+        "spo2": 97,
+        "rr": 22,
+        "temp": 37.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 36,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "2 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I saw a car crash happen right in front of me!",
+      "vitals": {
+        "hr": 115,
+        "bp_sys": 90,
+        "bp_dia": 55,
+        "spo2": 93,
+        "rr": 30,
+        "temp": 36.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 40,
+        "gender": "female",
+        "relevant_PMH": "anxiety",
+        "time_course": "15 minutes ago"
+      }
+    },
+    {
+      "chief_complaint": "My child is burning up and hardly responding.",
+      "vitals": {
+        "hr": 150,
+        "bp_sys": 85,
+        "bp_dia": 50,
+        "spo2": 92,
+        "rr": 36,
+        "temp": 41.0,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 5,
+        "gender": "male",
+        "relevant_PMH": "asthma",
+        "time_course": "started this morning"
+      }
+    },
+    {
+      "chief_complaint": "I was playing football and heard a snap in my back.",
+      "vitals": {
+        "hr": 100,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 98,
+        "rr": 20,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 16,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "1 hour ago"
+      }
+    },
+    {
+      "chief_complaint": "I had a fit and everything went blank for a while.",
+      "vitals": {
+        "hr": 98,
+        "bp_sys": 130,
+        "bp_dia": 85,
+        "spo2": 95,
+        "rr": 18,
+        "temp": 37.1,
+        "avpu": "P"
+      },
+      "history": {
+        "age": 46,
+        "gender": "female",
+        "relevant_PMH": "epilepsy",
+        "time_course": "stopped about 5 minutes ago"
+      }
+    }
+  ],
+  "3": [
+    {
+      "chief_complaint": "I can't catch my breath, it's been getting worse since yesterday.",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 135,
+        "bp_dia": 85,
+        "spo2": 88,
+        "rr": 26,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 67,
+        "gender": "male",
+        "relevant_PMH": "COPD",
+        "time_course": "Symptoms worsening over 24 hours"
+      }
+    },
+    {
+      "chief_complaint": "My leg is really swollen and it hurts to walk.",
+      "vitals": {
+        "hr": 90,
+        "bp_sys": 120,
+        "bp_dia": 80,
+        "spo2": 96,
+        "rr": 18,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 54,
+        "gender": "female",
+        "relevant_PMH": "None",
+        "time_course": "Swelling started two days ago"
+      }
+    },
+    {
+      "chief_complaint": "My back hurts so much, and I've been throwing up for hours.",
+      "vitals": {
+        "hr": 100,
+        "bp_sys": 125,
+        "bp_dia": 85,
+        "spo2": 97,
+        "rr": 20,
+        "temp": 38.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMH": "Kidney stones",
+        "time_course": "Pain started 8 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My ankle is swollen and it hurts a lot, but I don't remember falling.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 115,
+        "bp_dia": 75,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 30,
+        "gender": "female",
+        "relevant_PMH": "None",
+        "time_course": "Swelling and pain started this morning"
+      }
+    },
+    {
+      "chief_complaint": "It hurts to take a deep breath, and I feel out of breath.",
+      "vitals": {
+        "hr": 102,
+        "bp_sys": 122,
+        "bp_dia": 78,
+        "spo2": 94,
+        "rr": 22,
+        "temp": 37.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 40,
+        "gender": "male",
+        "relevant_PMH": "Asthma",
+        "time_course": "Symptoms started 12 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "Her leg is red and painful, she's been running a fever.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 115,
+        "bp_dia": 75,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 38.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 60,
+        "gender": "female",
+        "relevant_PMH": "Diabetes",
+        "time_course": "Leg redness and fever started yesterday"
+      }
+    },
+    {
+      "chief_complaint": "I keep throwing up and I have this awful pain in my side.",
+      "vitals": {
+        "hr": 105,
+        "bp_sys": 130,
+        "bp_dia": 85,
+        "spo2": 98,
+        "rr": 20,
+        "temp": 37.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 32,
+        "gender": "female",
+        "relevant_PMH": "None",
+        "time_course": "Symptoms started 10 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I'm feeling really confused, I can't remember much.",
+      "vitals": {
+        "hr": 82,
+        "bp_sys": 130,
+        "bp_dia": 80,
+        "spo2": 96,
+        "rr": 18,
+        "temp": 37.2,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 82,
+        "gender": "male",
+        "relevant_PMH": "Hypertension",
+        "time_course": "Confusion began 3 days ago"
+      }
+    },
+    {
+      "chief_complaint": "My tummy really hurts and I have a fever.",
+      "vitals": {
+        "hr": 112,
+        "bp_sys": 118,
+        "bp_dia": 76,
+        "spo2": 97,
+        "rr": 22,
+        "temp": 38.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 36,
+        "gender": "female",
+        "relevant_PMH": "Recurrent UTIs",
+        "time_course": "Pain started yesterday"
+      }
+    },
+    {
+      "chief_complaint": "The left side of my back is killing me. It hurts to move.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 120,
+        "bp_dia": 80,
+        "spo2": 99,
+        "rr": 18,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 50,
+        "gender": "female",
+        "relevant_PMH": "None",
+        "time_course": "Pain started suddenly 5 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My back hurts so bad, and I've been throwing up since yesterday.",
+      "vitals": {
+        "hr": 112,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 96,
+        "rr": 24,
+        "temp": 38.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 32,
+        "gender": "female",
+        "relevant_PMH": "none",
+        "time_course": "pain and vomiting started 12 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My leg is really swollen and hurts when I move.",
+      "vitals": {
+        "hr": 89,
+        "bp_sys": 132,
+        "bp_dia": 84,
+        "spo2": 98,
+        "rr": 17,
+        "temp": 37.3,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 56,
+        "gender": "male",
+        "relevant_PMH": "history of varicose veins",
+        "time_course": "swelling started 2 days ago"
+      }
+    },
+    {
+      "chief_complaint": "I've been coughing and struggling to breathe more than usual.",
+      "vitals": {
+        "hr": 108,
+        "bp_sys": 142,
+        "bp_dia": 88,
+        "spo2": 92,
+        "rr": 26,
+        "temp": 37.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 70,
+        "gender": "female",
+        "relevant_PMH": "COPD",
+        "time_course": "worsening over the last 3 days"
+      }
+    },
+    {
+      "chief_complaint": "I keep forgetting things and feel really confused.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 130,
+        "bp_dia": 75,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 82,
+        "gender": "male",
+        "relevant_PMH": "hypertension, mild cognitive impairment",
+        "time_course": "confusion worsened over 24 hours"
+      }
+    },
+    {
+      "chief_complaint": "My side hurts a lot, and I can't keep anything down.",
+      "vitals": {
+        "hr": 115,
+        "bp_sys": 105,
+        "bp_dia": 65,
+        "spo2": 95,
+        "rr": 20,
+        "temp": 37.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMH": "kidney stones",
+        "time_course": "pain started suddenly 10 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My leg is hurting and red, and I'm feeling feverish.",
+      "vitals": {
+        "hr": 102,
+        "bp_sys": 118,
+        "bp_dia": 78,
+        "spo2": 98,
+        "rr": 18,
+        "temp": 38.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 39,
+        "gender": "female",
+        "relevant_PMH": "diabetes",
+        "time_course": "redness and pain for 3 days"
+      }
+    },
+    {
+      "chief_complaint": "I'm wheezing a lot, and my inhaler isn't helping.",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 126,
+        "bp_dia": 82,
+        "spo2": 89,
+        "rr": 28,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 25,
+        "gender": "female",
+        "relevant_PMH": "asthma",
+        "time_course": "symptoms worsened in last 8 hours"
+      }
+    },
+    {
+      "chief_complaint": "My back really hurts, and I'm peeing a lot.",
+      "vitals": {
+        "hr": 105,
+        "bp_sys": 115,
+        "bp_dia": 70,
+        "spo2": 96,
+        "rr": 20,
+        "temp": 38.4,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 48,
+        "gender": "male",
+        "relevant_PMH": "recurrent UTIs",
+        "time_course": "symptoms started 24 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My ankle hurts so much, and it's hard to walk on it.",
+      "vitals": {
+        "hr": 98,
+        "bp_sys": 132,
+        "bp_dia": 84,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "injury happened while playing football 4 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My chest hurts when I take a deep breath.",
+      "vitals": {
+        "hr": 104,
+        "bp_sys": 124,
+        "bp_dia": 80,
+        "spo2": 95,
+        "rr": 22,
+        "temp": 37.4,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 52,
+        "gender": "female",
+        "relevant_PMH": "smoker",
+        "time_course": "started this morning"
+      }
+    },
+    {
+      "chief_complaint": "I've been wheezing and can't catch my breath sometimes.",
+      "vitals": {
+        "hr": 110,
+        "bp_sys": 140,
+        "bp_dia": 90,
+        "spo2": 91,
+        "rr": 28,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 65,
+        "gender": "male",
+        "relevant_pmh": "COPD",
+        "time_course": "2 days"
+      }
+    },
+    {
+      "chief_complaint": "My leg is swollen and hurts when I walk.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 125,
+        "bp_dia": 80,
+        "spo2": 98,
+        "rr": 20,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "female",
+        "relevant_pmh": null,
+        "time_course": "3 days"
+      }
+    },
+    {
+      "chief_complaint": "My lower back hurts and I feel sick to my stomach.",
+      "vitals": {
+        "hr": 100,
+        "bp_sys": 130,
+        "bp_dia": 85,
+        "spo2": 97,
+        "rr": 22,
+        "temp": 38.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 34,
+        "gender": "male",
+        "relevant_pmh": "History of kidney stones",
+        "time_course": "12 hours"
+      }
+    },
+    {
+      "chief_complaint": "I've been feeling weak and confused since last night.",
+      "vitals": {
+        "hr": 82,
+        "bp_sys": 115,
+        "bp_dia": 75,
+        "spo2": 96,
+        "rr": 18,
+        "temp": 37.2,
+        "avpu": "V"
+      },
+      "history": {
+        "age": 78,
+        "gender": "female",
+        "relevant_pmh": "Hypertension",
+        "time_course": "8 hours"
+      }
+    },
+    {
+      "chief_complaint": "I can't keep anything down, I keep throwing up.",
+      "vitals": {
+        "hr": 120,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 99,
+        "rr": 18,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 22,
+        "gender": "male",
+        "relevant_pmh": null,
+        "time_course": "24 hours"
+      }
+    },
+    {
+      "chief_complaint": "My hand is swollen and red, it's really sore.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 120,
+        "bp_dia": 80,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 37.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 55,
+        "gender": "female",
+        "relevant_pmh": "Diabetes",
+        "time_course": "2 days"
+      }
+    },
+    {
+      "chief_complaint": "It hurts on the side when I breathe in.",
+      "vitals": {
+        "hr": 105,
+        "bp_sys": 130,
+        "bp_dia": 85,
+        "spo2": 94,
+        "rr": 24,
+        "temp": 37.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 48,
+        "gender": "male",
+        "relevant_pmh": "Smoker",
+        "time_course": "1 day"
+      }
+    },
+    {
+      "chief_complaint": "My urine is super cloudy and my back hurts a lot.",
+      "vitals": {
+        "hr": 115,
+        "bp_sys": 125,
+        "bp_dia": 78,
+        "spo2": 97,
+        "rr": 20,
+        "temp": 38.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 32,
+        "gender": "female",
+        "relevant_pmh": "Frequent UTIs",
+        "time_course": "18 hours"
+      }
+    },
+    {
+      "chief_complaint": "I fell and now my arm hurts bad, it's swollen too.",
+      "vitals": {
+        "hr": 90,
+        "bp_sys": 135,
+        "bp_dia": 88,
+        "spo2": 99,
+        "rr": 18,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "female",
+        "relevant_pmh": null,
+        "time_course": "3 hours"
+      }
+    },
+    {
+      "chief_complaint": "I'm having a hard time breathing and my chest feels tight.",
+      "vitals": {
+        "hr": 130,
+        "bp_sys": 140,
+        "bp_dia": 92,
+        "spo2": 92,
+        "rr": 30,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 18,
+        "gender": "male",
+        "relevant_pmh": "Asthma",
+        "time_course": "6 hours"
+      }
+    }
+  ],
+  "4": [
+    {
+      "chief_complaint": "My wrist's been hurting since yesterday after I tripped. It's not too bad, but still hurts.",
+      "vitals": {
+        "hr": 82,
+        "bp_sys": 128,
+        "bp_dia": 75,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "female",
+        "relevant PMH": "none",
+        "time course": "pain started 24 hours ago after fall"
+      }
+    },
+    {
+      "chief_complaint": "I've got this itchy rash on my ankle after being outdoors. It's been a day.",
+      "vitals": {
+        "hr": 75,
+        "bp_sys": 120,
+        "bp_dia": 80,
+        "spo2": 99,
+        "rr": 15,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 41,
+        "gender": "male",
+        "relevant PMH": "seasonal allergies",
+        "time course": "itchy rash noticed 24 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I've been peeing a lot more than usual and it kind of stings. It's been happening for two days.",
+      "vitals": {
+        "hr": 78,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 98,
+        "rr": 18,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 52,
+        "gender": "female",
+        "relevant PMH": "recurrent UTIs",
+        "time course": "symptoms started 2 days ago"
+      }
+    },
+    {
+      "chief_complaint": "My ear's been aching and it's hard to hear out of it since yesterday morning.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 125,
+        "bp_dia": 78,
+        "spo2": 99,
+        "rr": 17,
+        "temp": 37.3,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 6,
+        "gender": "male",
+        "relevant PMH": "frequent ear infections",
+        "time course": "symptoms started yesterday morning"
+      }
+    },
+    {
+      "chief_complaint": "I have this weird bump on my eyelid for the last three days. It's kind of irritated.",
+      "vitals": {
+        "hr": 70,
+        "bp_sys": 115,
+        "bp_dia": 72,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 24,
+        "gender": "female",
+        "relevant PMH": "none",
+        "time course": "noticed the bump 3 days ago"
+      }
+    },
+    {
+      "chief_complaint": "The back of my hand got hit during a game and it's been bruised and sore for two days now.",
+      "vitals": {
+        "hr": 83,
+        "bp_sys": 130,
+        "bp_dia": 80,
+        "spo2": 99,
+        "rr": 15,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 34,
+        "gender": "male",
+        "relevant PMH": "none",
+        "time course": "injury occurred 2 days ago"
+      }
+    },
+    {
+      "chief_complaint": "My daughter's had this cough and a bit of a fever for the past day. She's still playing, though.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 105,
+        "bp_dia": 65,
+        "spo2": 97,
+        "rr": 20,
+        "temp": 38.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 3,
+        "gender": "female",
+        "relevant PMH": "none",
+        "time course": "symptoms began 24 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I got a small cut on my finger two days ago, and it's a little red and sore now.",
+      "vitals": {
+        "hr": 76,
+        "bp_sys": 118,
+        "bp_dia": 74,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 37.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 47,
+        "gender": "male",
+        "relevant PMH": "type 2 diabetes",
+        "time course": "injury occurred 2 days ago"
+      }
+    },
+    {
+      "chief_complaint": "I've had a headache that's lingered for a few days now. It's not super painful, but annoying.",
+      "vitals": {
+        "hr": 73,
+        "bp_sys": 115,
+        "bp_dia": 78,
+        "spo2": 99,
+        "rr": 14,
+        "temp": 36.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 58,
+        "gender": "female",
+        "relevant PMH": "chronic migraines",
+        "time course": "headache ongoing for 3 days"
+      }
+    },
+    {
+      "chief_complaint": "I sprained my ankle a couple of days ago, and it's swollen and a bit stiff.",
+      "vitals": {
+        "hr": 79,
+        "bp_sys": 122,
+        "bp_dia": 76,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 22,
+        "gender": "non-binary",
+        "relevant PMH": "none",
+        "time course": "injury occurred 2 days ago"
+      }
+    },
+    {
+      "chief_complaint": "My son has been complaining of ear pain since yesterday.",
+      "vitals": {
+        "hr": 86,
+        "bp_sys": 108,
+        "bp_dia": 72,
+        "spo2": 98,
+        "rr": 18,
+        "temp": 37.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 5,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "24 hours"
+      }
+    },
+    {
+      "chief_complaint": "I've had this low back ache for two days, especially when I bend over.",
+      "vitals": {
+        "hr": 78,
+        "bp_sys": 124,
+        "bp_dia": 80,
+        "spo2": 99,
+        "rr": 16,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 34,
+        "gender": "female",
+        "relevant_PMH": "none",
+        "time_course": "2 days"
+      }
+    },
+    {
+      "chief_complaint": "I tripped and scraped my knee, it's pretty sore and red now.",
+      "vitals": {
+        "hr": 92,
+        "bp_sys": 118,
+        "bp_dia": 78,
+        "spo2": 97,
+        "rr": 17,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 28,
+        "gender": "non-binary",
+        "relevant_PMH": "none",
+        "time_course": "12 hours"
+      }
+    },
+    {
+      "chief_complaint": "I have a sore throat and it's been feeling swollen since last night.",
+      "vitals": {
+        "hr": 80,
+        "bp_sys": 120,
+        "bp_dia": 78,
+        "spo2": 98,
+        "rr": 18,
+        "temp": 38.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 19,
+        "gender": "female",
+        "relevant_PMH": "tonsillitis",
+        "time_course": "1 night"
+      }
+    },
+    {
+      "chief_complaint": "I got stung by a bee and my arm is swollen but I feel fine.",
+      "vitals": {
+        "hr": 84,
+        "bp_sys": 122,
+        "bp_dia": 76,
+        "spo2": 99,
+        "rr": 16,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 42,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "6 hours"
+      }
+    },
+    {
+      "chief_complaint": "I've had this burning when I pee for two days, it's getting annoying.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 118,
+        "bp_dia": 74,
+        "spo2": 98,
+        "rr": 17,
+        "temp": 37.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 37,
+        "gender": "female",
+        "relevant_PMH": "recurrent UTIs",
+        "time_course": "2 days"
+      }
+    },
+    {
+      "chief_complaint": "My ankle's been hurting since I twisted it yesterday but I can walk.",
+      "vitals": {
+        "hr": 82,
+        "bp_sys": 126,
+        "bp_dia": 82,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "male",
+        "relevant_PMH": "none",
+        "time_course": "1 day"
+      }
+    },
+    {
+      "chief_complaint": "I think I have a cold, my nose is runny and I feel a bit tired.",
+      "vitals": {
+        "hr": 74,
+        "bp_sys": 115,
+        "bp_dia": 76,
+        "spo2": 99,
+        "rr": 18,
+        "temp": 37.4,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 25,
+        "gender": "female",
+        "relevant_PMH": "none",
+        "time_course": "3 days"
+      }
+    },
+    {
+      "chief_complaint": "I have a headache and my sinus feels stuffed up since this morning.",
+      "vitals": {
+        "hr": 76,
+        "bp_sys": 122,
+        "bp_dia": 80,
+        "spo2": 99,
+        "rr": 16,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_PMH": "sinusitis",
+        "time_course": "since morning"
+      }
+    },
+    {
+      "chief_complaint": "I've had this itchy rash on my arm for a few days, it's not going away.",
+      "vitals": {
+        "hr": 72,
+        "bp_sys": 120,
+        "bp_dia": 76,
+        "spo2": 98,
+        "rr": 17,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 32,
+        "gender": "female",
+        "relevant_PMH": "eczema",
+        "time_course": "few days"
+      }
+    },
+    {
+      "chief_complaint": "My ankle's a bit sore, I twisted it stepping off the curb yesterday.",
+      "vitals": {
+        "hr": 82,
+        "bp_sys": 118,
+        "bp_dia": 76,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 34,
+        "gender": "female",
+        "relevant_pmh": "None",
+        "time_course": "Pain started 24 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I've got a runny nose and scratchy throat since last night.",
+      "vitals": {
+        "hr": 78,
+        "bp_sys": 120,
+        "bp_dia": 80,
+        "spo2": 99,
+        "rr": 18,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "male",
+        "relevant_pmh": "Seasonal allergies",
+        "time_course": "Symptoms started 12 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I think I have a UTI again. It's uncomfortable when I pee.",
+      "vitals": {
+        "hr": 85,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 97,
+        "rr": 16,
+        "temp": 37.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 52,
+        "gender": "female",
+        "relevant_pmh": "Recurrent UTIs",
+        "time_course": "Symptoms began 2 days ago"
+      }
+    },
+    {
+      "chief_complaint": "My daughter has been pulling at her ear and crying more than usual.",
+      "vitals": {
+        "hr": 90,
+        "bp_sys": 104,
+        "bp_dia": 68,
+        "spo2": 99,
+        "rr": 20,
+        "temp": 38.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 3,
+        "gender": "female",
+        "relevant_pmh": "None",
+        "time_course": "Started last night"
+      }
+    },
+    {
+      "chief_complaint": "Got stung by a bee, it's a bit swollen but not too bad.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 115,
+        "bp_dia": 75,
+        "spo2": 98,
+        "rr": 17,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_pmh": "None",
+        "time_course": "Happened 3 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "My knee hurts from a soccer match yesterday, just a bit swollen.",
+      "vitals": {
+        "hr": 95,
+        "bp_sys": 112,
+        "bp_dia": 72,
+        "spo2": 98,
+        "rr": 15,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 17,
+        "gender": "male",
+        "relevant_pmh": "None",
+        "time_course": "Pain started 18 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "Cut my finger while chopping vegetables, it's not deep but still bleeds a bit.",
+      "vitals": {
+        "hr": 80,
+        "bp_sys": 120,
+        "bp_dia": 78,
+        "spo2": 99,
+        "rr": 16,
+        "temp": 36.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 28,
+        "gender": "female",
+        "relevant_pmh": "None",
+        "time_course": "Injury occurred 2 hours ago"
+      }
+    },
+    {
+      "chief_complaint": "I've had this cough and stuffy nose for a couple of days.",
+      "vitals": {
+        "hr": 76,
+        "bp_sys": 118,
+        "bp_dia": 79,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 37.3,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 40,
+        "gender": "female",
+        "relevant_pmh": "Asthma, controlled",
+        "time_course": "Symptoms started 3 days ago"
+      }
+    },
+    {
+      "chief_complaint": "I've been getting these headaches off and on, but they're not too bad.",
+      "vitals": {
+        "hr": 82,
+        "bp_sys": 130,
+        "bp_dia": 85,
+        "spo2": 98,
+        "rr": 14,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 48,
+        "gender": "male",
+        "relevant_pmh": "Hypertension",
+        "time_course": "Headaches started a week ago"
+      }
+    },
+    {
+      "chief_complaint": "My wrist is swollen, I think I sprained it when I fell off my bike yesterday.",
+      "vitals": {
+        "hr": 88,
+        "bp_sys": 116,
+        "bp_dia": 74,
+        "spo2": 99,
+        "rr": 16,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 23,
+        "gender": "female",
+        "relevant_pmh": "None",
+        "time_course": "Injury happened 24 hours ago"
+      }
+    }
+  ],
+  "5": [
+    {
+      "chief_complaint": "I need a refill on my blood pressure medicine.",
+      "vitals": {
+        "hr": 78,
+        "bp_sys": 135,
+        "bp_dia": 85,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 62,
+        "gender": "female",
+        "relevant_pmh": "hypertension",
+        "time_course": "ongoing for years"
+      }
+    },
+    {
+      "chief_complaint": "My back's been aching for a few weeks now, it's not getting worse but it's annoying.",
+      "vitals": {
+        "hr": 72,
+        "bp_sys": 120,
+        "bp_dia": 78,
+        "spo2": 99,
+        "rr": 14,
+        "temp": 36.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 45,
+        "gender": "male",
+        "relevant_pmh": "none",
+        "time_course": "3 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I've got this itchy rash that's not going away.",
+      "vitals": {
+        "hr": 76,
+        "bp_sys": 118,
+        "bp_dia": 76,
+        "spo2": 98,
+        "rr": 15,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 34,
+        "gender": "non-binary",
+        "relevant_pmh": "eczema",
+        "time_course": "1 week"
+      }
+    },
+    {
+      "chief_complaint": "I came for my diabetes medication review.",
+      "vitals": {
+        "hr": 80,
+        "bp_sys": 130,
+        "bp_dia": 82,
+        "spo2": 97,
+        "rr": 17,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 58,
+        "gender": "female",
+        "relevant_pmh": "type 2 diabetes",
+        "time_course": "stable"
+      }
+    },
+    {
+      "chief_complaint": "I've had a runny nose and a slight cold for a couple of weeks.",
+      "vitals": {
+        "hr": 68,
+        "bp_sys": 115,
+        "bp_dia": 75,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 37.2,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 28,
+        "gender": "male",
+        "relevant_pmh": "none",
+        "time_course": "2 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I've been having headaches every now and then, nothing too bad.",
+      "vitals": {
+        "hr": 74,
+        "bp_sys": 122,
+        "bp_dia": 80,
+        "spo2": 98,
+        "rr": 15,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 39,
+        "gender": "female",
+        "relevant_pmh": "migraines",
+        "time_course": "intermittent over 3 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I've been feeling a bit bloated, it's been a while now.",
+      "vitals": {
+        "hr": 70,
+        "bp_sys": 118,
+        "bp_dia": 79,
+        "spo2": 99,
+        "rr": 14,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 50,
+        "gender": "male",
+        "relevant_pmh": "irritable bowel syndrome",
+        "time_course": "4 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I need a note for work because I've been home with a cold.",
+      "vitals": {
+        "hr": 72,
+        "bp_sys": 120,
+        "bp_dia": 80,
+        "spo2": 97,
+        "rr": 18,
+        "temp": 37.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 31,
+        "gender": "female",
+        "relevant_pmh": "none",
+        "time_course": "5 days"
+      }
+    },
+    {
+      "chief_complaint": "I have these mild muscle cramps that come and go.",
+      "vitals": {
+        "hr": 75,
+        "bp_sys": 121,
+        "bp_dia": 79,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 47,
+        "gender": "non-binary",
+        "relevant_pmh": "slight electrolyte imbalance",
+        "time_course": "intermittent for 2 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I'm here for my annual asthma check-up.",
+      "vitals": {
+        "hr": 70,
+        "bp_sys": 119,
+        "bp_dia": 78,
+        "spo2": 96,
+        "rr": 17,
+        "temp": 36.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 22,
+        "gender": "male",
+        "relevant_pmh": "asthma",
+        "time_course": "control is good"
+      }
+    },
+    {
+      "chief_complaint": "I've run out of my blood pressure pills, can I get a new prescription?",
+      "vitals": {
+        "hr": 72,
+        "bp_sys": 130,
+        "bp_dia": 85,
+        "spo2": 98,
+        "rr": 16,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 67,
+        "gender": "male",
+        "relevant PMH": "hypertension",
+        "time course": "last refill was 3 months ago"
+      }
+    },
+    {
+      "chief_complaint": "I've got this itchy patch on my arm, it's been there for a while now.",
+      "vitals": {
+        "hr": 78,
+        "bp_sys": 122,
+        "bp_dia": 80,
+        "spo2": 98,
+        "rr": 14,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 22,
+        "gender": "female",
+        "relevant PMH": "eczema",
+        "time course": "3 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I need a sick note for work, I've had this back ache on and off for a month.",
+      "vitals": {
+        "hr": 70,
+        "bp_sys": 128,
+        "bp_dia": 82,
+        "spo2": 99,
+        "rr": 15,
+        "temp": 37.0,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 36,
+        "gender": "male",
+        "relevant PMH": "recurrent low back pain",
+        "time course": "1 month"
+      }
+    },
+    {
+      "chief_complaint": "I'm here for my regular diabetes medication check-up.",
+      "vitals": {
+        "hr": 74,
+        "bp_sys": 125,
+        "bp_dia": 78,
+        "spo2": 97,
+        "rr": 16,
+        "temp": 36.7,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 58,
+        "gender": "female",
+        "relevant PMH": "type 2 diabetes",
+        "time course": "routine check-up"
+      }
+    },
+    {
+      "chief_complaint": "My knee has been aching for weeks, should I be worried?",
+      "vitals": {
+        "hr": 72,
+        "bp_sys": 120,
+        "bp_dia": 76,
+        "spo2": 98,
+        "rr": 15,
+        "temp": 36.9,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 44,
+        "gender": "female",
+        "relevant PMH": "none",
+        "time course": "6 weeks"
+      }
+    },
+    {
+      "chief_complaint": "I need to get my asthma inhaler refilled, I ran out last week.",
+      "vitals": {
+        "hr": 80,
+        "bp_sys": 118,
+        "bp_dia": 75,
+        "spo2": 99,
+        "rr": 18,
+        "temp": 36.6,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 29,
+        "gender": "male",
+        "relevant PMH": "asthma",
+        "time course": "inhaler finished last week"
+      }
+    },
+    {
+      "chief_complaint": "I've had a headache, just dull, for about two weeks now.",
+      "vitals": {
+        "hr": 68,
+        "bp_sys": 115,
+        "bp_dia": 78,
+        "spo2": 98,
+        "rr": 14,
+        "temp": 37.1,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 40,
+        "gender": "non-binary",
+        "relevant PMH": "migraine",
+        "time course": "2 weeks"
+      }
+    },
+    {
+      "chief_complaint": "Can I get something for these occasional heartburn episodes?",
+      "vitals": {
+        "hr": 76,
+        "bp_sys": 120,
+        "bp_dia": 77,
+        "spo2": 97,
+        "rr": 16,
+        "temp": 36.5,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 55,
+        "gender": "female",
+        "relevant PMH": "none",
+        "time course": "over the past month"
+      }
+    },
+    {
+      "chief_complaint": "This cough won't go away, it's been annoying for about a month.",
+      "vitals": {
+        "hr": 72,
+        "bp_sys": 118,
+        "bp_dia": 75,
+        "spo2": 96,
+        "rr": 17,
+        "temp": 36.8,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 48,
+        "gender": "male",
+        "relevant PMH": "seasonal allergies",
+        "time course": "1 month"
+      }
+    },
+    {
+      "chief_complaint": "Just wanted to see if there's anything I can do for this dry skin, it's been like this for ages.",
+      "vitals": {
+        "hr": 68,
+        "bp_sys": 110,
+        "bp_dia": 70,
+        "spo2": 99,
+        "rr": 15,
+        "temp": 36.4,
+        "avpu": "A"
+      },
+      "history": {
+        "age": 61,
+        "gender": "female",
+        "relevant PMH": "hypothyroidism",
+        "time course": "several months"
+      }
+    }
+  ]
+}

data/gpt_train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1ac7b5b7901ba36504cc8ed7ca102b3b7e29bd5fac83225fdfd977ce7ff9f04
+size 126843

data/gpt_train_1768262967.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43d5315e895395ea1527ba8a9c7ce4b1be5ce0980e32e7072a61aed1cf2b4089
+size 126440

data/gpt_train_1768263238.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59614711514581ebb39937cc2de39b17f26385d0a3a8b609f8bfbe88ee45e747
+size 119683

data/gpt_train_1768263486.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c23626a2d5066662063c823e5ecf73d452ebed6a8ab16a126a001d823973087
+size 127940

data/gpt_train_1768263718.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:258cc804387a1ee0ec0aa2cb4aebe264417f354312e2ae74a1278f12ddf6463f
+size 127083

data/gpt_train_1768263967.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35fcb23391df7fb23f62fd363fd97981b7cd46c0cfe758546ac09595d3640f23
+size 127273

data/gpt_train_batch2.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19862a3186d9a1205ba62d11aeed4990b98e53ca04ddbe70ed694bf84c76a269
+size 127088

data/gpt_train_batch3.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dae84d7dcd8b6af4cd13175c0f6954028dfd819786d8557b24b80b5f3a0b5e63
+size 127042

data/gpt_train_batch4.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a65fd188d583f72e0048b04d9c428b60a35778d9f6c0f9eab3c07c17a5e890bd
+size 127784

data/gpt_train_batch5.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b8dbd36f3499ef0ee3421f8257ea667934b123d126837b2eda2bc5c6389e952
+size 127646

data/gpt_train_batch6.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1ac7b5b7901ba36504cc8ed7ca102b3b7e29bd5fac83225fdfd977ce7ff9f04
+size 126843

data/train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2190ce6670716d02d73bbab104ce778329e01c0f9457df4ead1266be403fbdb
+size 391884

data/train_expanded.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83f8d6551e957ba67bf874998d0026d761b7f8b892e59ab8521fce7193bc54d0
+size 1656706

data/val.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0da84966544b66084b27166bcbaa52ca15e5625c1ef89708713b3edd4d71937a
+size 78769

demo_human_play.py

@@ -0,0 +1,86 @@
+"""
+Demo script: Play the Triage Environment as a Human
+
+Run this to test the environment interactively.
+"""
+
+import sys
+sys.path.insert(0, '.')
+
+from nursesim_rl import TriageEnv
+
+
+def main():
+    env = TriageEnv(render_mode="human", seed=42)
+    obs, info = env.reset()
+    
+    print("\n🏥 Welcome to the A&E Triage Simulator!")
+    print("You are the Triage Nurse. Assess each patient and assign a category.\n")
+    
+    total_reward = 0
+    step = 0
+    
+    while True:
+        # Render current patient
+        env.render()
+        
+        if obs["patient_id"] == "":
+            print("\n✅ Shift complete! No more patients.")
+            break
+        
+        # Get user input
+        try:
+            category = int(input("\nEnter triage category (1-5): "))
+            if category < 1 or category > 5:
+                print("Invalid category. Please enter 1-5.")
+                continue
+        except ValueError:
+            print("Invalid input. Please enter a number.")
+            continue
+        
+        print("\nInterventions:")
+        for i, intervention in enumerate(env.INTERVENTIONS):
+            print(f"  [{i}] {intervention}")
+        
+        try:
+            intervention_idx = int(input("Choose intervention (0-6): "))
+            if intervention_idx < 0 or intervention_idx >= len(env.INTERVENTIONS):
+                intervention_idx = 0
+        except ValueError:
+            intervention_idx = 0
+        
+        # Take action
+        action = {
+            "triage_category": category,
+            "intervention": intervention_idx,
+        }
+        
+        obs, reward, terminated, truncated, info = env.step(action)
+        total_reward += reward
+        step += 1
+        
+        # Feedback
+        true_cat = info.get("true_category")
+        if true_cat and category == true_cat:
+            print(f"\n✅ Correct! Category {category} was right. Reward: +{reward:.1f}")
+        elif true_cat:
+            print(f"\n⚠️  The correct category was {true_cat}. You chose {category}. Reward: {reward:.1f}")
+        
+        if terminated or truncated:
+            break
+    
+    # Final stats
+    print("\n" + "="*60)
+    print("📊 SHIFT SUMMARY")
+    print("="*60)
+    print(f"  Patients Seen: {info.get('patients_seen', step)}")
+    print(f"  Correct Triage: {info.get('correct_triage', 0)}")
+    print(f"  Safety Failures: {info.get('safety_failures', 0)}")
+    print(f"  Total Reward: {total_reward:.1f}")
+    print("="*60)
+    
+    env.close()
+
+
+if __name__ == "__main__":
+    main()

generate_dataset.py

@@ -0,0 +1,185 @@
+"""
+Training Dataset Generator for NurseSim-RL
+
+Generates a dataset of triage scenarios with expert decisions for SFT training.
+Output format: JSONL compatible with Unsloth/TRL.
+"""
+
+import json
+import random
+from typing import Dict, List
+from pathlib import Path
+
+# Import from our environment
+import sys
+sys.path.insert(0, str(Path(__file__).parent))
+from nursesim_rl.patient_generator import PatientGenerator, SCENARIOS
+
+
+def format_observation(patient_data: Dict) -> str:
+    """Format patient data as a text observation for the LLM."""
+    vitals = patient_data["vitals"]
+    return f"""PATIENT PRESENTING TO A&E TRIAGE
+
+Chief Complaint: "{patient_data['complaint']}"
+
+Vitals:
+- HR: {vitals['hr']:.0f} bpm
+- BP: {vitals['bp_sys']:.0f}/{vitals['bp_dia']:.0f} mmHg
+- SpO2: {vitals['spo2']:.0f}%
+- RR: {vitals['rr']:.0f} /min
+- Temp: {vitals['temp']:.1f}C
+- AVPU: {vitals['avpu']}
+
+History: {patient_data['history']}
+
+WAITING ROOM: 12 patients | AVAILABLE BEDS: 4
+
+What is your triage decision?"""
+
+
+def get_expert_decision(category: int) -> Dict:
+    """Get the expert triage decision based on category."""
+    decisions = {
+        1: {
+            "category": 1,
+            "category_name": "Immediate (Red)",
+            "intervention": "send_to_resus",
+            "reasoning": "Life-threatening presentation requiring immediate resuscitation. Activate trauma/medical emergency team."
+        },
+        2: {
+            "category": 2,
+            "category_name": "Very Urgent (Orange)",
+            "intervention": "send_to_majors",
+            "reasoning": "Time-critical condition. Requires senior review within 10 minutes. Prioritise assessment."
+        },
+        3: {
+            "category": 3,
+            "category_name": "Urgent (Yellow)",
+            "intervention": "send_to_majors",
+            "reasoning": "Urgent presentation requiring assessment within 60 minutes. Monitor for deterioration."
+        },
+        4: {
+            "category": 4,
+            "category_name": "Standard (Green)",
+            "intervention": "send_to_minors",
+            "reasoning": "Stable presentation suitable for minor injuries/illness stream. Can wait safely."
+        },
+        5: {
+            "category": 5,
+            "category_name": "Non-urgent (Blue)",
+            "intervention": "refer_to_gp",
+            "reasoning": "Non-urgent presentation. Redirect to primary care or self-care advice."
+        },
+    }
+    return decisions[category]
+
+
+def format_response(decision: Dict) -> str:
+    """Format the expert decision as an LLM response."""
+    return f"""TRIAGE DECISION:
+
+Category: {decision['category']} - {decision['category_name']}
+Intervention: {decision['intervention']}
+
+Clinical Reasoning: {decision['reasoning']}"""
+
+
+def generate_dataset(n_samples: int = 500, seed: int = 42) -> List[Dict]:
+    """Generate a training dataset of triage scenarios."""
+    random.seed(seed)
+    dataset = []
+    
+    # Distribution adjusted for balanced IMMEDIATE case training
+    # Original: {1: 0.05, 2: 0.15, 3: 0.35, 4: 0.35, 5: 0.10}
+    category_weights = {1: 0.20, 2: 0.20, 3: 0.25, 4: 0.25, 5: 0.10}
+    
+    for i in range(n_samples):
+        # Weighted category selection
+        category = random.choices(
+            list(category_weights.keys()),
+            weights=list(category_weights.values())
+        )[0]
+        
+        # Get a random scenario for this category
+        scenario = random.choice(SCENARIOS[category])
+        
+        # Add some noise to vitals
+        noisy_vitals = {}
+        for k, v in scenario["vitals"].items():
+            if isinstance(v, (int, float)) and k != "avpu":
+                noise = random.gauss(0, abs(v) * 0.05) if v != 0 else 0
+                noisy_vitals[k] = v + noise
+            else:
+                noisy_vitals[k] = v
+        
+        patient_data = {
+            "complaint": scenario["chief_complaint"],
+            "vitals": noisy_vitals,
+            "history": scenario["history"],
+        }
+        
+        # Format as instruction-following example
+        observation = format_observation(patient_data)
+        decision = get_expert_decision(category)
+        response = format_response(decision)
+        
+        # Alpaca/ChatML format
+        example = {
+            "instruction": "You are an expert A&E Triage Nurse using the Manchester Triage System. Assess the following patient and provide your triage decision with clinical reasoning.",
+            "input": observation,
+            "output": response,
+            "category": category,  # For analysis
+        }
+        
+        dataset.append(example)
+    
+    return dataset
+
+
+def save_dataset(dataset: List[Dict], output_path: str):
+    """Save dataset to JSONL format."""
+    with open(output_path, 'w', encoding='utf-8') as f:
+        for example in dataset:
+            f.write(json.dumps(example, ensure_ascii=False) + '\n')
+    print(f"[OK] Saved {len(dataset)} examples to {output_path}")
+
+
+def main():
+    print("\n" + "="*60)
+    print("[DATASET] NurseSim-RL Training Data Generator")
+    print("="*60 + "\n")
+    
+    # Generate training set
+    print("Generating training dataset (500 examples)...")
+    train_data = generate_dataset(n_samples=500, seed=42)
+    save_dataset(train_data, "data/train.jsonl")
+    
+    # Generate validation set
+    print("Generating validation dataset (100 examples)...")
+    val_data = generate_dataset(n_samples=100, seed=123)
+    save_dataset(val_data, "data/val.jsonl")
+    
+    # Stats
+    print("\n" + "-"*40)
+    print("Dataset Statistics:")
+    for cat in range(1, 6):
+        train_count = sum(1 for x in train_data if x["category"] == cat)
+        val_count = sum(1 for x in val_data if x["category"] == cat)
+        print(f"  Category {cat}: {train_count} train / {val_count} val")
+    print("-"*40 + "\n")
+    
+    # Preview
+    print("Sample training example:")
+    print("-"*40)
+    sample = train_data[0]
+    print(f"[INSTRUCTION]\n{sample['instruction']}\n")
+    print(f"[INPUT]\n{sample['input']}\n")
+    print(f"[OUTPUT]\n{sample['output']}")
+    print("-"*40 + "\n")
+
+
+if __name__ == "__main__":
+    # Create data directory
+    Path("data").mkdir(exist_ok=True)
+    main()

generate_gpt_scenarios.py

@@ -0,0 +1,230 @@
+"""
+GPT-Powered Triage Scenario Generator
+Generates gold-standard synthetic training data using GPT-5/4
+"""
+
+import os
+import json
+import openai
+from pathlib import Path
+
+# Try to get API key from environment
+API_KEY = os.environ.get('OPENAI_API_KEY')
+if not API_KEY:
+    print("⚠️ Please set OPENAI_API_KEY environment variable")
+    print("Run: $env:OPENAI_API_KEY='your-key-here'")
+    exit(1)
+
+client = openai.OpenAI(api_key=API_KEY)
+
+# Find working model
+MODELS = ['gpt-4o', 'gpt-4-turbo', 'gpt-4', 'gpt-3.5-turbo']
+MODEL = None
+
+print("🔍 Finding available model...")
+for m in MODELS:
+    try:
+        client.chat.completions.create(model=m, messages=[{"role": "user", "content": "test"}], max_tokens=5)
+        MODEL = m
+        print(f"✅ Using: {m}")
+        break
+    except Exception as e:
+        continue
+
+if not MODEL:
+    print("❌ No model available")
+    exit(1)
+
+# System prompt for clinical accuracy
+SYSTEM_PROMPT = """You are an expert A&E Triage Nurse and clinical educator with 20 years of experience.
+You are creating realistic training scenarios for the Manchester Triage System.
+
+For each scenario, provide a JSON object with:
+1. chief_complaint: What the patient/relative actually SAYS (patient language, not medical jargon)
+2. vitals: Object with hr, bp_sys, bp_dia, spo2, rr, temp, avpu (A/V/P/U)
+3. history: Brief clinical history (age, gender, relevant PMH, time course)
+
+CRITICAL RULES:
+- Vitals MUST be physiologically consistent (shock = low BP + high HR)
+- Chief complaints use patient language, not medical terminology
+- Include diverse demographics
+- Cover atypical presentations
+
+Return a JSON array of 10 scenarios. No markdown, just JSON."""
+
+CATEGORY_PROMPTS = {
+    1: """Generate 10 Category 1 (IMMEDIATE/Red) triage scenarios.
+MTS Discriminators: Airway compromise, Inadequate breathing, Shock, Unresponsive, Currently fitting
+
+Include: STEMI, stroke, anaphylaxis, trauma, septic shock, cardiac arrest, status epilepticus, 
+massive GI bleed, tension pneumothorax, DKA coma, eclampsia, meningococcal sepsis, drowning, burns with inhalation.""",
+    
+    2: """Generate 10 Category 2 (VERY URGENT/Orange) scenarios.
+MTS Discriminators: Severe pain, Altered consciousness, Very hot adult/child, Significant mechanism
+
+Include: Chest pain (possible ACS), thunderclap headache, focal neurology, high fever with red flags, 
+significant trauma, acute abdomen, spinal injury, ongoing seizure that stopped.""",
+    
+    3: """Generate 10 Category 3 (URGENT/Yellow) scenarios.
+MTS Discriminators: Moderate pain, Hot adult/child, Persistent vomiting, Pleuritic pain
+
+Include: COPD exacerbation, cellulitis, renal colic, fractures, moderate asthma, 
+acute confusion (elderly), DVT symptoms, pyelonephritis.""",
+    
+    4: """Generate 10 Category 4 (STANDARD/Green) scenarios.
+MTS Discriminators: Recent mild pain, Warm, Recent problem
+
+Include: Minor injuries, viral illnesses, stable chronic conditions, minor lacerations, 
+sprains, insect bites, UTI symptoms, ear infections.""",
+    
+    5: """Generate 10 Category 5 (NON-URGENT/Blue) scenarios.
+MTS Discriminators: Recent mild problem
+
+Include: Prescription requests, chronic stable issues, minor rashes, medication reviews, 
+sick notes, minor aches lasting weeks."""
+}
+
+
+def generate_batch(category: int, batch_num: int) -> list:
+    """Generate a batch of scenarios for a category"""
+    try:
+        response = client.chat.completions.create(
+            model=MODEL,
+            messages=[
+                {"role": "system", "content": SYSTEM_PROMPT},
+                {"role": "user", "content": CATEGORY_PROMPTS[category] + f"\n\nBatch {batch_num} - ensure unique scenarios not seen before."}
+            ],
+            temperature=0.9,
+            max_tokens=4000
+        )
+        
+        content = response.choices[0].message.content.strip()
+        # Clean markdown
+        if content.startswith('```'):
+            content = content.split('```')[1]
+            if content.startswith('json'):
+                content = content[4:]
+        content = content.strip()
+        
+        try:
+            return json.loads(content)
+        except json.JSONDecodeError:
+            print(f"    ⚠️ JSON Error. Attempting repair...")
+            # Simple repair: try to find the list bracket
+            start = content.find('[')
+            end = content.rfind(']')
+            if start != -1 and end != -1:
+                try:
+                    return json.loads(content[start:end+1])
+                except:
+                    pass
+            print(f"    ❌ Failed to parse JSON batch.")
+            return []
+            
+    except Exception as e:
+        print(f"    Error: {str(e)[:60]}")
+        return []
+
+
+def format_training_example(scenario: dict, category: int) -> dict:
+    """Convert scenario to training format"""
+    vitals = scenario.get('vitals', {})
+    
+    input_text = f"""PATIENT PRESENTING TO A&E TRIAGE
+
+Chief Complaint: "{scenario.get('chief_complaint', 'Unknown')}"
+
+Vitals:
+- HR: {vitals.get('hr', 80):.0f} bpm
+- BP: {vitals.get('bp_sys', 120):.0f}/{vitals.get('bp_dia', 80):.0f} mmHg
+- SpO2: {vitals.get('spo2', 98):.0f}%
+- RR: {vitals.get('rr', 16):.0f} /min
+- Temp: {vitals.get('temp', 37.0):.1f}C
+- AVPU: {vitals.get('avpu', 'A')}
+
+History: {scenario.get('history', 'Unknown')}
+
+WAITING ROOM: 12 patients | AVAILABLE BEDS: 4
+
+What is your triage decision?"""
+
+    decisions = {
+        1: ("Immediate (Red)", "send_to_resus", "Life-threatening presentation requiring immediate resuscitation."),
+        2: ("Very Urgent (Orange)", "send_to_majors", "Time-critical condition. Requires senior review within 10 minutes."),
+        3: ("Urgent (Yellow)", "send_to_majors", "Urgent presentation requiring assessment within 60 minutes."),
+        4: ("Standard (Green)", "send_to_minors", "Stable presentation suitable for minor injuries/illness stream."),
+        5: ("Non-urgent (Blue)", "refer_to_gp", "Non-urgent presentation. Redirect to primary care.")
+    }
+    
+    name, action, reason = decisions[category]
+    output_text = f"""TRIAGE DECISION:
+
+Category: {category} - {name}
+Intervention: {action}
+
+Clinical Reasoning: {reason}"""
+
+    return {
+        "instruction": "You are an expert A&E Triage Nurse using the Manchester Triage System. Assess the following patient and provide your triage decision with clinical reasoning.",
+        "input": input_text,
+        "output": output_text,
+        "category": category
+    }
+
+
+def main():
+    print("\n" + "="*60)
+    print("🏥 GPT-Powered Triage Scenario Generator")
+    print("="*60 + "\n")
+    
+    # Configure batches per category
+    batches = {1: 5, 2: 3, 3: 3, 4: 3, 5: 2}
+    
+    all_scenarios = {}
+    training_data = []
+    
+    for cat in [1, 2, 3, 4, 5]:
+        print(f"\n📋 Category {cat} ({batches[cat]} batches):")
+        all_scenarios[cat] = []
+        
+        for b in range(batches[cat]):
+            print(f"  Batch {b+1}...", end=" ")
+            scenarios = generate_batch(cat, b+1)
+            all_scenarios[cat].extend(scenarios)
+            print(f"✅ {len(scenarios)} scenarios")
+        
+        # Convert to training format
+        for s in all_scenarios[cat]:
+            try:
+                example = format_training_example(s, cat)
+                training_data.append(example)
+            except:
+                continue
+        
+        print(f"  Total Cat {cat}: {len(all_scenarios[cat])} scenarios")
+    
+    # Save raw scenarios
+    with open('data/gpt_scenarios.json', 'w') as f:
+        json.dump(all_scenarios, f, indent=2)
+    print(f"\n✅ Saved raw scenarios: data/gpt_scenarios.json")
+    
+    # Save as JSONL with unique timestamp
+    import time
+    timestamp = int(time.time())
+    output_file = f'data/gpt_train_{timestamp}.jsonl'
+    
+    with open(output_file, 'w') as f:
+        for ex in training_data:
+            f.write(json.dumps(ex) + '\n')
+    print(f"✅ Saved training data: {output_file} ({len(training_data)} examples)")
+    
+    # skipped automatic merging to avoid race conditions
+    print(f"✅ Saved batch: {output_file} ({len(training_data)} examples)")
+    
+    print("\n" + "="*60)
+    print("✅ Generation complete!")
+    print("="*60 + "\n")
+
+
+if __name__ == "__main__":
+    main()

nursesim_rl/__init__.py

@@ -1,10 +1,16 @@
-"""
-NurseSim-RL: A Triage Environment for Reinforcement Learning
-OpenEnv Challenge Entry - 2026
-"""
-
-from .triage_env import TriageEnv
-from .patient_generator import PatientGenerator
-
-__version__ = "0.1.0"
-__all__ = ["TriageEnv", "PatientGenerator"]
+"""
+NurseSim-RL: A Triage Environment for Reinforcement Learning
+OpenEnv Challenge Entry - 2026
+"""
+
+from .triage_env import TriageEnv
+from .patient_generator import PatientGenerator
+from .semantic_wrapper import NurseEmbedWrapper, make_semantic_triage_env
+
+__version__ = "0.2.0"
+__all__ = [
+    "TriageEnv", 
+    "PatientGenerator",
+    "NurseEmbedWrapper",
+    "make_semantic_triage_env",
+]

nursesim_rl/patient_generator.py

@@ -1,272 +1,272 @@
-"""
-Patient Generator for NurseSim-RL
-
-Generates synthetic patient scenarios based on Manchester Triage System categories.
-"""
-
-import random
-from dataclasses import dataclass
-from typing import Dict, List, Optional
-
-
-@dataclass
-class Patient:
-    """Represents a patient presenting to A&E."""
-    id: str
-    chief_complaint: str
-    vitals: Dict[str, float]
-    history: str
-    true_category: int  # 1-5 (Ground truth for reward calculation)
-    time_arrived: int
-    
-
-# Manchester Triage System Scenarios
-# Gold-standard scenarios validated against MTS discriminators and clinical guidelines
-SCENARIOS = {
-    # Category 1: Immediate (Red) - Life threatening
-    # MTS Discriminators: Airway compromise, Inadequate breathing, Shock, Unresponsive
-    1: [
-        # Original scenarios
-        {
-            "chief_complaint": "I can't breathe... my chest is crushing... the pain goes down my arm.",
-            "vitals": {"hr": 120, "bp_sys": 85, "bp_dia": 50, "spo2": 88, "rr": 32, "temp": 36.5, "avpu": "V"},
-            "history": "65yo male, known cardiac history, sudden onset 20 mins ago."
-        },
-        {
-            "chief_complaint": "He collapsed and isn't responding to me!",
-            "vitals": {"hr": 0, "bp_sys": 0, "bp_dia": 0, "spo2": 0, "rr": 0, "temp": 35.0, "avpu": "U"},
-            "history": "72yo male found unresponsive by wife. Bystander CPR in progress."
-        },
-        {
-            "chief_complaint": "My face is swelling up and I can't swallow... I ate shellfish.",
-            "vitals": {"hr": 130, "bp_sys": 70, "bp_dia": 40, "spo2": 85, "rr": 28, "temp": 37.0, "avpu": "A"},
-            "history": "28yo female, known shellfish allergy, stridor audible."
-        },
-        # NEW: STEMI - Classic presentation
-        {
-            "chief_complaint": "There's an elephant sitting on my chest... I feel like I'm going to die.",
-            "vitals": {"hr": 110, "bp_sys": 80, "bp_dia": 55, "spo2": 91, "rr": 28, "temp": 36.8, "avpu": "A"},
-            "history": "58yo male, sweating profusely, nausea, pain radiating to jaw and left arm, 30 mins onset."
-        },
-        # NEW: Acute Stroke - FAST positive
-        {
-            "chief_complaint": "My husband's face has dropped and he can't speak properly!",
-            "vitals": {"hr": 95, "bp_sys": 195, "bp_dia": 110, "spo2": 96, "rr": 18, "temp": 37.0, "avpu": "V"},
-            "history": "70yo male, sudden onset 45 mins ago, right-sided weakness, slurred speech, facial droop."
-        },
-        # NEW: Tension Pneumothorax
-        {
-            "chief_complaint": "I was stabbed... I can't breathe... everything is going dark.",
-            "vitals": {"hr": 135, "bp_sys": 75, "bp_dia": 45, "spo2": 78, "rr": 40, "temp": 36.5, "avpu": "V"},
-            "history": "25yo male, stab wound to right chest, trachea deviated left, absent breath sounds right."
-        },
-        # NEW: Septic Shock
-        {
-            "chief_complaint": "She's been confused all day and now she's going cold and clammy.",
-            "vitals": {"hr": 125, "bp_sys": 70, "bp_dia": 40, "spo2": 88, "rr": 30, "temp": 39.8, "avpu": "V"},
-            "history": "78yo female, recent UTI, mottled skin, delayed capillary refill, drowsy."
-        },
-        # NEW: Major Trauma - MVA
-        {
-            "chief_complaint": "He was thrown from the car... he's not making sense!",
-            "vitals": {"hr": 130, "bp_sys": 85, "bp_dia": 50, "spo2": 90, "rr": 32, "temp": 35.5, "avpu": "V"},
-            "history": "35yo male, unrestrained driver, high-speed RTC, GCS 9, deformed left femur."
-        },
-        # NEW: Massive GI Bleed
-        {
-            "chief_complaint": "I've vomited loads of blood and now I feel faint...",
-            "vitals": {"hr": 140, "bp_sys": 75, "bp_dia": 45, "spo2": 94, "rr": 26, "temp": 36.2, "avpu": "V"},
-            "history": "55yo male, known liver disease, haematemesis x3, melena, pale and diaphoretic."
-        },
-        # NEW: Status Epilepticus
-        {
-            "chief_complaint": "My daughter won't stop fitting! It's been going on for ages!",
-            "vitals": {"hr": 145, "bp_sys": 160, "bp_dia": 95, "spo2": 82, "rr": 8, "temp": 38.5, "avpu": "U"},
-            "history": "8yo female, known epilepsy, continuous tonic-clonic seizure for 12 minutes, cyanosed."
-        },
-        # NEW: DKA with Coma
-        {
-            "chief_complaint": "He's a diabetic and now he won't wake up properly.",
-            "vitals": {"hr": 115, "bp_sys": 90, "bp_dia": 55, "spo2": 95, "rr": 35, "temp": 37.2, "avpu": "P"},
-            "history": "22yo male, Type 1 diabetic, Kussmaul breathing, fruity breath, missed insulin for 3 days."
-        },
-        # NEW: Eclampsia
-        {
-            "chief_complaint": "She's 8 months pregnant and started fitting!",
-            "vitals": {"hr": 120, "bp_sys": 180, "bp_dia": 120, "spo2": 89, "rr": 24, "temp": 37.5, "avpu": "U"},
-            "history": "32yo female, 34 weeks pregnant, seizure witnessed, severe headache and visual disturbance earlier."
-        },
-        # NEW: Complete Airway Obstruction
-        {
-            "chief_complaint": "He was eating steak and now he can't breathe at all!",
-            "vitals": {"hr": 140, "bp_sys": 160, "bp_dia": 100, "spo2": 65, "rr": 0, "temp": 37.0, "avpu": "A"},
-            "history": "60yo male, choking at restaurant, cannot cough or speak, universal choking sign, cyanotic."
-        },
-        # NEW: Haemorrhagic Stroke (SAH)
-        {
-            "chief_complaint": "It's the worst headache of my life... like being hit with a hammer.",
-            "vitals": {"hr": 55, "bp_sys": 200, "bp_dia": 115, "spo2": 94, "rr": 14, "temp": 37.3, "avpu": "V"},
-            "history": "48yo female, sudden onset thunderclap headache, vomiting, photophobia, now drowsy."
-        },
-        # NEW: Ruptured AAA
-        {
-            "chief_complaint": "The pain in my back is unbearable... I feel like I'm going to pass out.",
-            "vitals": {"hr": 125, "bp_sys": 75, "bp_dia": 50, "spo2": 93, "rr": 28, "temp": 36.0, "avpu": "V"},
-            "history": "72yo male, known AAA, sudden severe abdominal/back pain, pulsatile mass, pale and sweating."
-        },
-        # NEW: Atypical ACS (Elderly)
-        {
-            "chief_complaint": "I just feel really unwell... something is very wrong.",
-            "vitals": {"hr": 45, "bp_sys": 80, "bp_dia": 50, "spo2": 88, "rr": 24, "temp": 36.5, "avpu": "V"},
-            "history": "82yo female, known diabetes, vague malaise, nausea, cool and clammy, no chest pain."
-        },
-        # NEW: Meningococcal Septicaemia
-        {
-            "chief_complaint": "My child has a rash that won't go away when I press it!",
-            "vitals": {"hr": 180, "bp_sys": 70, "bp_dia": 40, "spo2": 90, "rr": 45, "temp": 40.2, "avpu": "V"},
-            "history": "4yo male, non-blanching purpuric rash, headache, photophobia, neck stiffness, drowsy."
-        },
-        # NEW: Drowning/Near-drowning
-        {
-            "chief_complaint": "He was pulled from the pool and he's not breathing properly!",
-            "vitals": {"hr": 50, "bp_sys": 70, "bp_dia": 45, "spo2": 60, "rr": 6, "temp": 34.0, "avpu": "U"},
-            "history": "6yo male, submersion ~5 mins, bystander CPR given, now gasping, hypothermic."
-        },
-        # NEW: Severe Burns with Inhalation
-        {
-            "chief_complaint": "She was in a house fire... her voice sounds strange now.",
-            "vitals": {"hr": 130, "bp_sys": 90, "bp_dia": 55, "spo2": 85, "rr": 32, "temp": 37.5, "avpu": "A"},
-            "history": "45yo female, facial burns, singed nasal hairs, hoarse voice, stridor developing."
-        },
-    ],
-    
-    # Category 2: Very Urgent (Orange) - Time critical
-    2: [
-        {
-            "chief_complaint": "I have the worst headache of my life. It came on suddenly.",
-            "vitals": {"hr": 90, "bp_sys": 180, "bp_dia": 100, "spo2": 97, "rr": 18, "temp": 37.2, "avpu": "A"},
-            "history": "45yo female, sudden onset occipital headache, photophobia, neck stiffness."
-        },
-        {
-            "chief_complaint": "My little boy is having a fit and won't stop!",
-            "vitals": {"hr": 150, "bp_sys": 90, "bp_dia": 55, "spo2": 90, "rr": 24, "temp": 39.5, "avpu": "U"},
-            "history": "3yo male, febrile seizure ongoing for 8 minutes."
-        },
-        {
-            "chief_complaint": "I fell and I can't feel my legs.",
-            "vitals": {"hr": 100, "bp_sys": 140, "bp_dia": 85, "spo2": 98, "rr": 20, "temp": 36.8, "avpu": "A"},
-            "history": "55yo male, fell from ladder, complaining of neck pain, no sensation below T4."
-        },
-    ],
-    
-    # Category 3: Urgent (Yellow)
-    3: [
-        {
-            "chief_complaint": "I've had abdominal pain for 2 days. It's getting worse and I'm vomiting.",
-            "vitals": {"hr": 105, "bp_sys": 110, "bp_dia": 70, "spo2": 97, "rr": 20, "temp": 38.2, "avpu": "A"},
-            "history": "32yo female, RIF pain, guarding, rebound tenderness."
-        },
-        {
-            "chief_complaint": "I've been short of breath for a few days. It's worse when I walk.",
-            "vitals": {"hr": 95, "bp_sys": 125, "bp_dia": 80, "spo2": 92, "rr": 24, "temp": 37.0, "avpu": "A"},
-            "history": "70yo male, COPD, productive cough, increased work of breathing."
-        },
-        {
-            "chief_complaint": "I cut my hand on a knife. It won't stop bleeding.",
-            "vitals": {"hr": 88, "bp_sys": 130, "bp_dia": 82, "spo2": 99, "rr": 16, "temp": 36.9, "avpu": "A"},
-            "history": "40yo male, deep laceration to palm, tendon visible, bleeding controlled with pressure."
-        },
-    ],
-    
-    # Category 4: Standard (Green)
-    4: [
-        {
-            "chief_complaint": "I've had a sore throat and cough for 3 days.",
-            "vitals": {"hr": 78, "bp_sys": 120, "bp_dia": 75, "spo2": 99, "rr": 14, "temp": 37.8, "avpu": "A"},
-            "history": "25yo female, coryzal symptoms, no difficulty swallowing, eating and drinking well."
-        },
-        {
-            "chief_complaint": "I twisted my ankle playing football yesterday.",
-            "vitals": {"hr": 72, "bp_sys": 118, "bp_dia": 72, "spo2": 99, "rr": 14, "temp": 36.8, "avpu": "A"},
-            "history": "22yo male, swollen lateral ankle, can weight bear with pain, no deformity."
-        },
-        {
-            "chief_complaint": "I've had diarrhoea and vomiting since last night.",
-            "vitals": {"hr": 85, "bp_sys": 115, "bp_dia": 70, "spo2": 98, "rr": 16, "temp": 37.5, "avpu": "A"},
-            "history": "35yo female, kept down fluids this morning, passing urine, no blood in stool."
-        },
-    ],
-    
-    # Category 5: Non-urgent (Blue)
-    5: [
-        {
-            "chief_complaint": "I need a repeat prescription for my blood pressure tablets.",
-            "vitals": {"hr": 70, "bp_sys": 135, "bp_dia": 85, "spo2": 99, "rr": 14, "temp": 36.7, "avpu": "A"},
-            "history": "60yo male, ran out of Amlodipine, asymptomatic."
-        },
-        {
-            "chief_complaint": "I've had a rash on my arm for a week. It's itchy.",
-            "vitals": {"hr": 68, "bp_sys": 120, "bp_dia": 78, "spo2": 99, "rr": 14, "temp": 36.8, "avpu": "A"},
-            "history": "30yo female, localised erythematous rash, no systemic symptoms, not spreading."
-        },
-        {
-            "chief_complaint": "I just want my sick note signing.",
-            "vitals": {"hr": 72, "bp_sys": 122, "bp_dia": 80, "spo2": 99, "rr": 14, "temp": 36.8, "avpu": "A"},
-            "history": "45yo male, recovering from back strain, no red flags."
-        },
-    ],
-}
-
-
-class PatientGenerator:
-    """Generates patient scenarios for the Triage environment."""
-    
-    def __init__(self, seed: Optional[int] = None):
-        if seed is not None:
-            random.seed(seed)
-        self._patient_count = 0
-    
-    def generate(self, category: Optional[int] = None) -> Patient:
-        """
-        Generate a random patient.
-        
-        Args:
-            category: Optional specific category (1-5). If None, weighted random selection.
-        
-        Returns:
-            A Patient object.
-        """
-        if category is None:
-            # Weighted distribution mimicking real A&E (more Cat 3-4 than Cat 1)
-            weights = [5, 15, 35, 35, 10]  # % distribution
-            category = random.choices([1, 2, 3, 4, 5], weights=weights)[0]
-        
-        scenario = random.choice(SCENARIOS[category])
-        self._patient_count += 1
-        
-        # Add some noise to vitals
-        noisy_vitals = {
-            k: v + random.gauss(0, v * 0.05) if isinstance(v, float) else v
-            for k, v in scenario["vitals"].items()
-        }
-        
-        return Patient(
-            id=f"P{self._patient_count:04d}",
-            chief_complaint=scenario["chief_complaint"],
-            vitals=noisy_vitals,
-            history=scenario["history"],
-            true_category=category,
-            time_arrived=0,  # Will be set by environment
-        )
-    
-    def generate_batch(self, n: int) -> List[Patient]:
-        """Generate a batch of n patients."""
-        return [self.generate() for _ in range(n)]
-
-
-if __name__ == "__main__":
-    # Quick test
-    gen = PatientGenerator(seed=42)
-    for _ in range(5):
-        patient = gen.generate()
-        print(f"{patient.id}: Cat {patient.true_category} - {patient.chief_complaint[:50]}...")
+"""
+Patient Generator for NurseSim-RL
+
+Generates synthetic patient scenarios based on Manchester Triage System categories.
+"""
+
+import random
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+
+@dataclass
+class Patient:
+    """Represents a patient presenting to A&E."""
+    id: str
+    chief_complaint: str
+    vitals: Dict[str, float]
+    history: str
+    true_category: int  # 1-5 (Ground truth for reward calculation)
+    time_arrived: int
+    
+
+# Manchester Triage System Scenarios
+# Gold-standard scenarios validated against MTS discriminators and clinical guidelines
+SCENARIOS = {
+    # Category 1: Immediate (Red) - Life threatening
+    # MTS Discriminators: Airway compromise, Inadequate breathing, Shock, Unresponsive
+    1: [
+        # Original scenarios
+        {
+            "chief_complaint": "I can't breathe... my chest is crushing... the pain goes down my arm.",
+            "vitals": {"hr": 120, "bp_sys": 85, "bp_dia": 50, "spo2": 88, "rr": 32, "temp": 36.5, "avpu": "V"},
+            "history": "65yo male, known cardiac history, sudden onset 20 mins ago."
+        },
+        {
+            "chief_complaint": "He collapsed and isn't responding to me!",
+            "vitals": {"hr": 0, "bp_sys": 0, "bp_dia": 0, "spo2": 0, "rr": 0, "temp": 35.0, "avpu": "U"},
+            "history": "72yo male found unresponsive by wife. Bystander CPR in progress."
+        },
+        {
+            "chief_complaint": "My face is swelling up and I can't swallow... I ate shellfish.",
+            "vitals": {"hr": 130, "bp_sys": 70, "bp_dia": 40, "spo2": 85, "rr": 28, "temp": 37.0, "avpu": "A"},
+            "history": "28yo female, known shellfish allergy, stridor audible."
+        },
+        # NEW: STEMI - Classic presentation
+        {
+            "chief_complaint": "There's an elephant sitting on my chest... I feel like I'm going to die.",
+            "vitals": {"hr": 110, "bp_sys": 80, "bp_dia": 55, "spo2": 91, "rr": 28, "temp": 36.8, "avpu": "A"},
+            "history": "58yo male, sweating profusely, nausea, pain radiating to jaw and left arm, 30 mins onset."
+        },
+        # NEW: Acute Stroke - FAST positive
+        {
+            "chief_complaint": "My husband's face has dropped and he can't speak properly!",
+            "vitals": {"hr": 95, "bp_sys": 195, "bp_dia": 110, "spo2": 96, "rr": 18, "temp": 37.0, "avpu": "V"},
+            "history": "70yo male, sudden onset 45 mins ago, right-sided weakness, slurred speech, facial droop."
+        },
+        # NEW: Tension Pneumothorax
+        {
+            "chief_complaint": "I was stabbed... I can't breathe... everything is going dark.",
+            "vitals": {"hr": 135, "bp_sys": 75, "bp_dia": 45, "spo2": 78, "rr": 40, "temp": 36.5, "avpu": "V"},
+            "history": "25yo male, stab wound to right chest, trachea deviated left, absent breath sounds right."
+        },
+        # NEW: Septic Shock
+        {
+            "chief_complaint": "She's been confused all day and now she's going cold and clammy.",
+            "vitals": {"hr": 125, "bp_sys": 70, "bp_dia": 40, "spo2": 88, "rr": 30, "temp": 39.8, "avpu": "V"},
+            "history": "78yo female, recent UTI, mottled skin, delayed capillary refill, drowsy."
+        },
+        # NEW: Major Trauma - MVA
+        {
+            "chief_complaint": "He was thrown from the car... he's not making sense!",
+            "vitals": {"hr": 130, "bp_sys": 85, "bp_dia": 50, "spo2": 90, "rr": 32, "temp": 35.5, "avpu": "V"},
+            "history": "35yo male, unrestrained driver, high-speed RTC, GCS 9, deformed left femur."
+        },
+        # NEW: Massive GI Bleed
+        {
+            "chief_complaint": "I've vomited loads of blood and now I feel faint...",
+            "vitals": {"hr": 140, "bp_sys": 75, "bp_dia": 45, "spo2": 94, "rr": 26, "temp": 36.2, "avpu": "V"},
+            "history": "55yo male, known liver disease, haematemesis x3, melena, pale and diaphoretic."
+        },
+        # NEW: Status Epilepticus
+        {
+            "chief_complaint": "My daughter won't stop fitting! It's been going on for ages!",
+            "vitals": {"hr": 145, "bp_sys": 160, "bp_dia": 95, "spo2": 82, "rr": 8, "temp": 38.5, "avpu": "U"},
+            "history": "8yo female, known epilepsy, continuous tonic-clonic seizure for 12 minutes, cyanosed."
+        },
+        # NEW: DKA with Coma
+        {
+            "chief_complaint": "He's a diabetic and now he won't wake up properly.",
+            "vitals": {"hr": 115, "bp_sys": 90, "bp_dia": 55, "spo2": 95, "rr": 35, "temp": 37.2, "avpu": "P"},
+            "history": "22yo male, Type 1 diabetic, Kussmaul breathing, fruity breath, missed insulin for 3 days."
+        },
+        # NEW: Eclampsia
+        {
+            "chief_complaint": "She's 8 months pregnant and started fitting!",
+            "vitals": {"hr": 120, "bp_sys": 180, "bp_dia": 120, "spo2": 89, "rr": 24, "temp": 37.5, "avpu": "U"},
+            "history": "32yo female, 34 weeks pregnant, seizure witnessed, severe headache and visual disturbance earlier."
+        },
+        # NEW: Complete Airway Obstruction
+        {
+            "chief_complaint": "He was eating steak and now he can't breathe at all!",
+            "vitals": {"hr": 140, "bp_sys": 160, "bp_dia": 100, "spo2": 65, "rr": 0, "temp": 37.0, "avpu": "A"},
+            "history": "60yo male, choking at restaurant, cannot cough or speak, universal choking sign, cyanotic."
+        },
+        # NEW: Haemorrhagic Stroke (SAH)
+        {
+            "chief_complaint": "It's the worst headache of my life... like being hit with a hammer.",
+            "vitals": {"hr": 55, "bp_sys": 200, "bp_dia": 115, "spo2": 94, "rr": 14, "temp": 37.3, "avpu": "V"},
+            "history": "48yo female, sudden onset thunderclap headache, vomiting, photophobia, now drowsy."
+        },
+        # NEW: Ruptured AAA
+        {
+            "chief_complaint": "The pain in my back is unbearable... I feel like I'm going to pass out.",
+            "vitals": {"hr": 125, "bp_sys": 75, "bp_dia": 50, "spo2": 93, "rr": 28, "temp": 36.0, "avpu": "V"},
+            "history": "72yo male, known AAA, sudden severe abdominal/back pain, pulsatile mass, pale and sweating."
+        },
+        # NEW: Atypical ACS (Elderly)
+        {
+            "chief_complaint": "I just feel really unwell... something is very wrong.",
+            "vitals": {"hr": 45, "bp_sys": 80, "bp_dia": 50, "spo2": 88, "rr": 24, "temp": 36.5, "avpu": "V"},
+            "history": "82yo female, known diabetes, vague malaise, nausea, cool and clammy, no chest pain."
+        },
+        # NEW: Meningococcal Septicaemia
+        {
+            "chief_complaint": "My child has a rash that won't go away when I press it!",
+            "vitals": {"hr": 180, "bp_sys": 70, "bp_dia": 40, "spo2": 90, "rr": 45, "temp": 40.2, "avpu": "V"},
+            "history": "4yo male, non-blanching purpuric rash, headache, photophobia, neck stiffness, drowsy."
+        },
+        # NEW: Drowning/Near-drowning
+        {
+            "chief_complaint": "He was pulled from the pool and he's not breathing properly!",
+            "vitals": {"hr": 50, "bp_sys": 70, "bp_dia": 45, "spo2": 60, "rr": 6, "temp": 34.0, "avpu": "U"},
+            "history": "6yo male, submersion ~5 mins, bystander CPR given, now gasping, hypothermic."
+        },
+        # NEW: Severe Burns with Inhalation
+        {
+            "chief_complaint": "She was in a house fire... her voice sounds strange now.",
+            "vitals": {"hr": 130, "bp_sys": 90, "bp_dia": 55, "spo2": 85, "rr": 32, "temp": 37.5, "avpu": "A"},
+            "history": "45yo female, facial burns, singed nasal hairs, hoarse voice, stridor developing."
+        },
+    ],
+    
+    # Category 2: Very Urgent (Orange) - Time critical
+    2: [
+        {
+            "chief_complaint": "I have the worst headache of my life. It came on suddenly.",
+            "vitals": {"hr": 90, "bp_sys": 180, "bp_dia": 100, "spo2": 97, "rr": 18, "temp": 37.2, "avpu": "A"},
+            "history": "45yo female, sudden onset occipital headache, photophobia, neck stiffness."
+        },
+        {
+            "chief_complaint": "My little boy is having a fit and won't stop!",
+            "vitals": {"hr": 150, "bp_sys": 90, "bp_dia": 55, "spo2": 90, "rr": 24, "temp": 39.5, "avpu": "U"},
+            "history": "3yo male, febrile seizure ongoing for 8 minutes."
+        },
+        {
+            "chief_complaint": "I fell and I can't feel my legs.",
+            "vitals": {"hr": 100, "bp_sys": 140, "bp_dia": 85, "spo2": 98, "rr": 20, "temp": 36.8, "avpu": "A"},
+            "history": "55yo male, fell from ladder, complaining of neck pain, no sensation below T4."
+        },
+    ],
+    
+    # Category 3: Urgent (Yellow)
+    3: [
+        {
+            "chief_complaint": "I've had abdominal pain for 2 days. It's getting worse and I'm vomiting.",
+            "vitals": {"hr": 105, "bp_sys": 110, "bp_dia": 70, "spo2": 97, "rr": 20, "temp": 38.2, "avpu": "A"},
+            "history": "32yo female, RIF pain, guarding, rebound tenderness."
+        },
+        {
+            "chief_complaint": "I've been short of breath for a few days. It's worse when I walk.",
+            "vitals": {"hr": 95, "bp_sys": 125, "bp_dia": 80, "spo2": 92, "rr": 24, "temp": 37.0, "avpu": "A"},
+            "history": "70yo male, COPD, productive cough, increased work of breathing."
+        },
+        {
+            "chief_complaint": "I cut my hand on a knife. It won't stop bleeding.",
+            "vitals": {"hr": 88, "bp_sys": 130, "bp_dia": 82, "spo2": 99, "rr": 16, "temp": 36.9, "avpu": "A"},
+            "history": "40yo male, deep laceration to palm, tendon visible, bleeding controlled with pressure."
+        },
+    ],
+    
+    # Category 4: Standard (Green)
+    4: [
+        {
+            "chief_complaint": "I've had a sore throat and cough for 3 days.",
+            "vitals": {"hr": 78, "bp_sys": 120, "bp_dia": 75, "spo2": 99, "rr": 14, "temp": 37.8, "avpu": "A"},
+            "history": "25yo female, coryzal symptoms, no difficulty swallowing, eating and drinking well."
+        },
+        {
+            "chief_complaint": "I twisted my ankle playing football yesterday.",
+            "vitals": {"hr": 72, "bp_sys": 118, "bp_dia": 72, "spo2": 99, "rr": 14, "temp": 36.8, "avpu": "A"},
+            "history": "22yo male, swollen lateral ankle, can weight bear with pain, no deformity."
+        },
+        {
+            "chief_complaint": "I've had diarrhoea and vomiting since last night.",
+            "vitals": {"hr": 85, "bp_sys": 115, "bp_dia": 70, "spo2": 98, "rr": 16, "temp": 37.5, "avpu": "A"},
+            "history": "35yo female, kept down fluids this morning, passing urine, no blood in stool."
+        },
+    ],
+    
+    # Category 5: Non-urgent (Blue)
+    5: [
+        {
+            "chief_complaint": "I need a repeat prescription for my blood pressure tablets.",
+            "vitals": {"hr": 70, "bp_sys": 135, "bp_dia": 85, "spo2": 99, "rr": 14, "temp": 36.7, "avpu": "A"},
+            "history": "60yo male, ran out of Amlodipine, asymptomatic."
+        },
+        {
+            "chief_complaint": "I've had a rash on my arm for a week. It's itchy.",
+            "vitals": {"hr": 68, "bp_sys": 120, "bp_dia": 78, "spo2": 99, "rr": 14, "temp": 36.8, "avpu": "A"},
+            "history": "30yo female, localised erythematous rash, no systemic symptoms, not spreading."
+        },
+        {
+            "chief_complaint": "I just want my sick note signing.",
+            "vitals": {"hr": 72, "bp_sys": 122, "bp_dia": 80, "spo2": 99, "rr": 14, "temp": 36.8, "avpu": "A"},
+            "history": "45yo male, recovering from back strain, no red flags."
+        },
+    ],
+}
+
+
+class PatientGenerator:
+    """Generates patient scenarios for the Triage environment."""
+    
+    def __init__(self, seed: Optional[int] = None):
+        if seed is not None:
+            random.seed(seed)
+        self._patient_count = 0
+    
+    def generate(self, category: Optional[int] = None) -> Patient:
+        """
+        Generate a random patient.
+        
+        Args:
+            category: Optional specific category (1-5). If None, weighted random selection.
+        
+        Returns:
+            A Patient object.
+        """
+        if category is None:
+            # Weighted distribution mimicking real A&E (more Cat 3-4 than Cat 1)
+            weights = [5, 15, 35, 35, 10]  # % distribution
+            category = random.choices([1, 2, 3, 4, 5], weights=weights)[0]
+        
+        scenario = random.choice(SCENARIOS[category])
+        self._patient_count += 1
+        
+        # Add some noise to vitals
+        noisy_vitals = {
+            k: v + random.gauss(0, v * 0.05) if isinstance(v, float) else v
+            for k, v in scenario["vitals"].items()
+        }
+        
+        return Patient(
+            id=f"P{self._patient_count:04d}",
+            chief_complaint=scenario["chief_complaint"],
+            vitals=noisy_vitals,
+            history=scenario["history"],
+            true_category=category,
+            time_arrived=0,  # Will be set by environment
+        )
+    
+    def generate_batch(self, n: int) -> List[Patient]:
+        """Generate a batch of n patients."""
+        return [self.generate() for _ in range(n)]
+
+
+if __name__ == "__main__":
+    # Quick test
+    gen = PatientGenerator(seed=42)
+    for _ in range(5):
+        patient = gen.generate()
+        print(f"{patient.id}: Cat {patient.true_category} - {patient.chief_complaint[:50]}...")

nursesim_rl/semantic_wrapper.py

@@ -0,0 +1,169 @@
+"""
+NurseEmbedWrapper: A Gymnasium wrapper that converts text observations to NurseEmbed vectors.
+
+This enables Language-Conditioned Reinforcement Learning for nursing scenarios.
+"""
+
+import gymnasium as gym
+from gymnasium import spaces
+import numpy as np
+from typing import Any, Dict, Tuple
+
+# Lazy load NurseEmbed to avoid import errors if not available
+_embed_model = None
+
+def _get_embed_model():
+    """Lazy load the NurseEmbed model."""
+    global _embed_model
+    if _embed_model is None:
+        try:
+            from sentence_transformers import SentenceTransformer
+            _embed_model = SentenceTransformer("NurseCitizenDeveloper/NurseEmbed-300M")
+            print("[OK] NurseEmbed model loaded successfully")
+        except Exception as e:
+            print(f"[WARN] NurseEmbed not available: {e}")
+            # Fallback to random embeddings for testing
+            _embed_model = "fallback"
+    return _embed_model
+
+
+class NurseEmbedWrapper(gym.Wrapper):
+    """
+    Wraps a TriageEnv and converts text observations to NurseEmbed vectors.
+    Also flattens the Dict action space to MultiDiscrete for SB3 compatibility.
+    
+    Instead of the agent seeing:
+        {"chief_complaint": "Chest pain radiating to arm...", "vitals": {...}, ...}
+    
+    It sees:
+        np.array([...390 dimensional semantic vector...])
+    
+    The vector encodes the MEANING of the clinical presentation.
+    """
+    
+    EMBED_DIM = 384  # NurseEmbed-300M outputs 384D vectors (nomic base)
+    
+    def __init__(self, env: gym.Env, use_vitals: bool = True):
+        """
+        Args:
+            env: The underlying TriageEnv
+            use_vitals: Whether to append vitals to the embedding
+        """
+        super().__init__(env)
+        
+        self.use_vitals = use_vitals
+        self.model = _get_embed_model()
+        
+        # Calculate observation dimension
+        obs_dim = self.EMBED_DIM
+        if use_vitals:
+            obs_dim += 6  # HR, BP_sys, BP_dia, SpO2, RR, Temp
+        
+        # Override observation space to be a flat Box
+        self.observation_space = spaces.Box(
+            low=-np.inf,
+            high=np.inf,
+            shape=(obs_dim,),
+            dtype=np.float32
+        )
+        
+        # Flatten action space for SB3 compatibility
+        # Original: Dict({'triage_category': Discrete(5, start=1), 'intervention': Discrete(7)})
+        # New: MultiDiscrete([5, 7]) where first dim is category (0-4, add 1 later) and second is intervention
+        self.action_space = spaces.MultiDiscrete([5, 7])
+        
+        # Cache for embeddings (same text -> same embedding)
+        self._embedding_cache: Dict[str, np.ndarray] = {}
+    
+    def reset(self, **kwargs) -> Tuple[np.ndarray, Dict]:
+        """Reset and convert observation."""
+        obs, info = self.env.reset(**kwargs)
+        return self._convert_observation(obs), info
+    
+    def step(self, action: np.ndarray) -> Tuple[np.ndarray, float, bool, bool, Dict]:
+        """Convert action, step, convert observation."""
+        # Convert flat action [category_idx, intervention_idx] to Dict
+        dict_action = {
+            "triage_category": int(action[0]) + 1,  # 0-4 -> 1-5
+            "intervention": int(action[1])
+        }
+        obs, reward, terminated, truncated, info = self.env.step(dict_action)
+        return self._convert_observation(obs), reward, terminated, truncated, info
+    
+    def _convert_observation(self, obs: Dict) -> np.ndarray:
+        """Convert Dict observation to semantic vector."""
+        # Build text representation
+        text = self._build_clinical_text(obs)
+        
+        # Get embedding (with caching)
+        embedding = self._get_embedding(text)
+        
+        # Optionally append vitals
+        if self.use_vitals:
+            vitals_vector = self._extract_vitals(obs)
+            embedding = np.concatenate([embedding, vitals_vector])
+        
+        return embedding.astype(np.float32)
+    
+    def _build_clinical_text(self, obs: Dict) -> str:
+        """Build a clinical description from the observation."""
+        complaint = obs.get("chief_complaint", "Unknown complaint")
+        history = obs.get("history", "")
+        
+        # Create a rich clinical description
+        text = f"Patient presents with: {complaint}. Clinical history: {history}."
+        
+        # Add vitals context as text for semantic understanding
+        vitals = obs.get("vitals", {})
+        if vitals:
+            text += f" Vital signs: HR {vitals.get('hr', 'N/A')}, "
+            text += f"BP {vitals.get('bp_sys', 'N/A')}/{vitals.get('bp_dia', 'N/A')}, "
+            text += f"SpO2 {vitals.get('spo2', 'N/A')}%, "
+            text += f"RR {vitals.get('rr', 'N/A')}, "
+            text += f"AVPU {vitals.get('avpu', 'A')}."
+        
+        return text
+    
+    def _get_embedding(self, text: str) -> np.ndarray:
+        """Get embedding with caching."""
+        if text in self._embedding_cache:
+            return self._embedding_cache[text]
+        
+        if self.model == "fallback":
+            # Fallback: deterministic pseudo-random embedding based on text hash
+            np.random.seed(hash(text) % 2**32)
+            embedding = np.random.randn(self.EMBED_DIM)
+        else:
+            embedding = self.model.encode(text, normalize_embeddings=True)
+        
+        self._embedding_cache[text] = embedding
+        return embedding
+    
+    def _extract_vitals(self, obs: Dict) -> np.ndarray:
+        """Extract vitals as a normalized vector."""
+        vitals = obs.get("vitals", {})
+        return np.array([
+            vitals.get("hr", 70) / 200.0,  # Normalize HR
+            vitals.get("bp_sys", 120) / 200.0,
+            vitals.get("bp_dia", 80) / 150.0,
+            vitals.get("spo2", 98) / 100.0,
+            vitals.get("rr", 16) / 40.0,
+            (vitals.get("temp", 37.0) - 35) / 5.0,  # Normalize temp
+        ], dtype=np.float32)
+
+
+def make_semantic_triage_env(seed: int = None, **kwargs) -> gym.Env:
+    """Factory function to create a semantically-aware triage environment."""
+    from nursesim_rl import TriageEnv
+    
+    base_env = TriageEnv(seed=seed, **kwargs)
+    wrapped_env = NurseEmbedWrapper(base_env, use_vitals=True)
+    
+    return wrapped_env
+
+
+# Register wrapped version
+gym.register(
+    id="NurseSim-SemanticTriage-v0",
+    entry_point="nursesim_rl.semantic_wrapper:make_semantic_triage_env",
+)

nursesim_rl/triage_env.py

@@ -1,303 +1,303 @@
-"""
-TriageEnv: A Gymnasium-compatible RL environment for A&E Triage.
-
-OpenEnv Challenge Entry - 2026
-"""
-
-import gymnasium as gym
-from gymnasium import spaces
-import numpy as np
-from typing import Any, Dict, Optional, Tuple
-
-from .patient_generator import PatientGenerator, Patient
-
-
-class TriageEnv(gym.Env):
-    """
-    A&E Triage Environment.
-    
-    The agent plays the role of a Triage Nurse, assessing patients and
-    assigning them to the correct Manchester Triage System category.
-    
-    Observation:
-        - patient_complaint (str): The patient's chief complaint
-        - vitals (dict): HR, BP, SpO2, RR, Temp, AVPU
-        - history (str): Brief clinical history
-        - waiting_room (int): Number of patients currently waiting
-        - available_beds (int): Beds available in Resus/Majors
-    
-    Action:
-        - triage_category (int): 1-5 (Immediate to Non-urgent)
-        - intervention (str): One of the allowed interventions
-    
-    Reward:
-        - +10 for correct triage category
-        - +5 for adjacent category (within 1)
-        - -50 for critical safety failure (under-triaging P1/P2 by 2+ levels)
-        - -1 per minute waiting for high-acuity patients
-    """
-    
-    metadata = {"render_modes": ["human", "ansi"], "render_fps": 1}
-    
-    INTERVENTIONS = [
-        "send_to_resus",
-        "send_to_majors", 
-        "send_to_minors",
-        "order_ecg",
-        "give_analgesia",
-        "discharge",
-        "refer_to_gp",
-    ]
-    
-    def __init__(
-        self,
-        max_patients: int = 20,
-        max_steps: int = 50,
-        render_mode: Optional[str] = None,
-        seed: Optional[int] = None,
-    ):
-        super().__init__()
-        
-        self.max_patients = max_patients
-        self.max_steps = max_steps
-        self.render_mode = render_mode
-        
-        self.patient_generator = PatientGenerator(seed=seed)
-        
-        # Action space: Discrete triage category + intervention
-        self.action_space = spaces.Dict({
-            "triage_category": spaces.Discrete(5, start=1),  # 1-5
-            "intervention": spaces.Discrete(len(self.INTERVENTIONS)),
-        })
-        
-        # Observation space
-        self.observation_space = spaces.Dict({
-            "patient_id": spaces.Text(10),
-            "chief_complaint": spaces.Text(500),
-            "vitals": spaces.Dict({
-                "hr": spaces.Box(0, 300, shape=(), dtype=np.float32),
-                "bp_sys": spaces.Box(0, 300, shape=(), dtype=np.float32),
-                "bp_dia": spaces.Box(0, 200, shape=(), dtype=np.float32),
-                "spo2": spaces.Box(0, 100, shape=(), dtype=np.float32),
-                "rr": spaces.Box(0, 60, shape=(), dtype=np.float32),
-                "temp": spaces.Box(30, 45, shape=(), dtype=np.float32),
-                "avpu": spaces.Text(1),
-            }),
-            "history": spaces.Text(500),
-            "waiting_room": spaces.Discrete(100),
-            "available_beds": spaces.Discrete(20),
-        })
-        
-        # State
-        self.current_patient: Optional[Patient] = None
-        self.waiting_queue: list = []
-        self.step_count: int = 0
-        self.total_reward: float = 0.0
-        self.available_beds: int = 10
-        self.episode_stats: Dict[str, Any] = {}
-        
-    def reset(
-        self,
-        seed: Optional[int] = None,
-        options: Optional[Dict] = None,
-    ) -> Tuple[Dict, Dict]:
-        """Reset the environment to initial state."""
-        super().reset(seed=seed)
-        
-        if seed is not None:
-            self.patient_generator = PatientGenerator(seed=seed)
-        
-        # Reset state
-        self.step_count = 0
-        self.total_reward = 0.0
-        self.available_beds = 10
-        self.episode_stats = {
-            "correct_triage": 0,
-            "safety_failures": 0,
-            "patients_seen": 0,
-        }
-        
-        # Generate initial waiting room
-        initial_patients = np.random.randint(3, 8)
-        self.waiting_queue = self.patient_generator.generate_batch(initial_patients)
-        for i, p in enumerate(self.waiting_queue):
-            p.time_arrived = -i * 5  # Stagger arrival times
-        
-        # Get first patient
-        self.current_patient = self._get_next_patient()
-        
-        return self._get_observation(), self._get_info()
-    
-    def step(self, action: Dict) -> Tuple[Dict, float, bool, bool, Dict]:
-        """
-        Execute one step in the environment.
-        
-        Args:
-            action: Dict with 'triage_category' (1-5) and 'intervention' (index)
-        
-        Returns:
-            observation, reward, terminated, truncated, info
-        """
-        self.step_count += 1
-        
-        if self.current_patient is None:
-            # No more patients - episode ends
-            return self._get_observation(), 0.0, True, False, self._get_info()
-        
-        # Parse action
-        assigned_category = action.get("triage_category", 3)
-        intervention_idx = action.get("intervention", 0)
-        intervention = self.INTERVENTIONS[intervention_idx]
-        
-        # Calculate reward
-        reward = self._calculate_reward(assigned_category, intervention)
-        self.total_reward += reward
-        self.episode_stats["patients_seen"] += 1
-        
-        # Update bed availability based on intervention
-        if intervention in ["send_to_resus", "send_to_majors"]:
-            self.available_beds = max(0, self.available_beds - 1)
-        elif intervention in ["discharge", "refer_to_gp"]:
-            self.available_beds = min(10, self.available_beds + 1)
-        
-        # Possibly add new patients to queue
-        if np.random.random() < 0.3:  # 30% chance of new arrival
-            new_patient = self.patient_generator.generate()
-            new_patient.time_arrived = self.step_count
-            self.waiting_queue.append(new_patient)
-        
-        # Get next patient
-        self.current_patient = self._get_next_patient()
-        
-        # Check termination
-        terminated = self.current_patient is None and len(self.waiting_queue) == 0
-        truncated = self.step_count >= self.max_steps
-        
-        return self._get_observation(), reward, terminated, truncated, self._get_info()
-    
-    def _calculate_reward(self, assigned_category: int, intervention: str) -> float:
-        """Calculate reward based on triage decision."""
-        if self.current_patient is None:
-            return 0.0
-        
-        true_category = self.current_patient.true_category
-        category_diff = abs(assigned_category - true_category)
-        
-        reward = 0.0
-        
-        # Category accuracy
-        if category_diff == 0:
-            reward += 10.0
-            self.episode_stats["correct_triage"] += 1
-        elif category_diff == 1:
-            reward += 5.0  # Close enough
-        else:
-            reward -= 5.0 * category_diff  # Penalty scales with error
-        
-        # Critical safety failure: Under-triaging a critical patient
-        if true_category <= 2 and assigned_category >= true_category + 2:
-            reward -= 50.0
-            self.episode_stats["safety_failures"] += 1
-        
-        # Intervention appropriateness
-        if true_category == 1 and intervention == "send_to_resus":
-            reward += 5.0
-        elif true_category == 5 and intervention in ["discharge", "refer_to_gp"]:
-            reward += 3.0
-        elif true_category == 1 and intervention == "discharge":
-            reward -= 30.0  # Never discharge a P1!
-        
-        return reward
-    
-    def _get_next_patient(self) -> Optional[Patient]:
-        """Get the next patient from the queue (FIFO with priority override)."""
-        if not self.waiting_queue:
-            return None
-        
-        # Priority override: P1 patients jump the queue
-        for i, patient in enumerate(self.waiting_queue):
-            if patient.true_category == 1:
-                return self.waiting_queue.pop(i)
-        
-        # Otherwise FIFO
-        return self.waiting_queue.pop(0)
-    
-    def _get_observation(self) -> Dict:
-        """Build the observation dictionary."""
-        if self.current_patient is None:
-            return {
-                "patient_id": "",
-                "chief_complaint": "No patients waiting.",
-                "vitals": {
-                    "hr": 0.0, "bp_sys": 0.0, "bp_dia": 0.0,
-                    "spo2": 0.0, "rr": 0.0, "temp": 0.0, "avpu": "A"
-                },
-                "history": "",
-                "waiting_room": len(self.waiting_queue),
-                "available_beds": self.available_beds,
-            }
-        
-        return {
-            "patient_id": self.current_patient.id,
-            "chief_complaint": self.current_patient.chief_complaint,
-            "vitals": {
-                "hr": float(self.current_patient.vitals.get("hr", 0)),
-                "bp_sys": float(self.current_patient.vitals.get("bp_sys", 0)),
-                "bp_dia": float(self.current_patient.vitals.get("bp_dia", 0)),
-                "spo2": float(self.current_patient.vitals.get("spo2", 0)),
-                "rr": float(self.current_patient.vitals.get("rr", 0)),
-                "temp": float(self.current_patient.vitals.get("temp", 0)),
-                "avpu": str(self.current_patient.vitals.get("avpu", "A")),
-            },
-            "history": self.current_patient.history,
-            "waiting_room": len(self.waiting_queue),
-            "available_beds": self.available_beds,
-        }
-    
-    def _get_info(self) -> Dict:
-        """Return additional info."""
-        return {
-            "step": self.step_count,
-            "total_reward": self.total_reward,
-            "true_category": self.current_patient.true_category if self.current_patient else None,
-            **self.episode_stats,
-        }
-    
-    def render(self) -> Optional[str]:
-        """Render the environment."""
-        if self.render_mode == "human" or self.render_mode == "ansi":
-            obs = self._get_observation()
-            output = f"""
-╔══════════════════════════════════════════════════════════════════╗
-║  A&E TRIAGE SIMULATOR  │  Step: {self.step_count:3d} │ Waiting: {obs['waiting_room']:2d} │ Beds: {obs['available_beds']:2d}  ║
-╠══════════════════════════════════════════════════════════════════╣
-║  PATIENT: {obs['patient_id']:<54} ║
-╠──────────────────────────────────────────────────────────────────╣
-║  Chief Complaint:                                                ║
-║    "{obs['chief_complaint'][:60]:<60}" ║
-╠──────────────────────────────────────────────────────────────────╣
-║  VITALS:                                                         ║
-║    HR: {obs['vitals']['hr']:>3.0f}  │  BP: {obs['vitals']['bp_sys']:>3.0f}/{obs['vitals']['bp_dia']:<3.0f}  │  SpO2: {obs['vitals']['spo2']:>3.0f}%            ║
-║    RR: {obs['vitals']['rr']:>3.0f}  │  Temp: {obs['vitals']['temp']:.1f}°C  │  AVPU: {obs['vitals']['avpu']}               ║
-╠──────────────────────────────────────────────────────────────────╣
-║  History: {obs['history'][:55]:<55} ║
-╠══════════════════════════════════════════════════════════════════╣
-║  What is your triage decision?                                   ║
-║    [1] Immediate  [2] Very Urgent  [3] Urgent  [4] Std  [5] Non  ║
-╚══════════════════════════════════════════════════════════════════╝
-"""
-            if self.render_mode == "human":
-                print(output)
-            return output
-        return None
-    
-    def close(self):
-        """Clean up resources."""
-        pass
-
-
-# Register with Gymnasium
-gym.register(
-    id="NurseSim-Triage-v0",
-    entry_point="nursesim_rl:TriageEnv",
-)
+"""
+TriageEnv: A Gymnasium-compatible RL environment for A&E Triage.
+
+OpenEnv Challenge Entry - 2026
+"""
+
+import gymnasium as gym
+from gymnasium import spaces
+import numpy as np
+from typing import Any, Dict, Optional, Tuple
+
+from .patient_generator import PatientGenerator, Patient
+
+
+class TriageEnv(gym.Env):
+    """
+    A&E Triage Environment.
+    
+    The agent plays the role of a Triage Nurse, assessing patients and
+    assigning them to the correct Manchester Triage System category.
+    
+    Observation:
+        - patient_complaint (str): The patient's chief complaint
+        - vitals (dict): HR, BP, SpO2, RR, Temp, AVPU
+        - history (str): Brief clinical history
+        - waiting_room (int): Number of patients currently waiting
+        - available_beds (int): Beds available in Resus/Majors
+    
+    Action:
+        - triage_category (int): 1-5 (Immediate to Non-urgent)
+        - intervention (str): One of the allowed interventions
+    
+    Reward:
+        - +10 for correct triage category
+        - +5 for adjacent category (within 1)
+        - -50 for critical safety failure (under-triaging P1/P2 by 2+ levels)
+        - -1 per minute waiting for high-acuity patients
+    """
+    
+    metadata = {"render_modes": ["human", "ansi"], "render_fps": 1}
+    
+    INTERVENTIONS = [
+        "send_to_resus",
+        "send_to_majors", 
+        "send_to_minors",
+        "order_ecg",
+        "give_analgesia",
+        "discharge",
+        "refer_to_gp",
+    ]
+    
+    def __init__(
+        self,
+        max_patients: int = 20,
+        max_steps: int = 50,
+        render_mode: Optional[str] = None,
+        seed: Optional[int] = None,
+    ):
+        super().__init__()
+        
+        self.max_patients = max_patients
+        self.max_steps = max_steps
+        self.render_mode = render_mode
+        
+        self.patient_generator = PatientGenerator(seed=seed)
+        
+        # Action space: Discrete triage category + intervention
+        self.action_space = spaces.Dict({
+            "triage_category": spaces.Discrete(5, start=1),  # 1-5
+            "intervention": spaces.Discrete(len(self.INTERVENTIONS)),
+        })
+        
+        # Observation space
+        self.observation_space = spaces.Dict({
+            "patient_id": spaces.Text(10),
+            "chief_complaint": spaces.Text(500),
+            "vitals": spaces.Dict({
+                "hr": spaces.Box(0, 300, shape=(), dtype=np.float32),
+                "bp_sys": spaces.Box(0, 300, shape=(), dtype=np.float32),
+                "bp_dia": spaces.Box(0, 200, shape=(), dtype=np.float32),
+                "spo2": spaces.Box(0, 100, shape=(), dtype=np.float32),
+                "rr": spaces.Box(0, 60, shape=(), dtype=np.float32),
+                "temp": spaces.Box(30, 45, shape=(), dtype=np.float32),
+                "avpu": spaces.Text(1),
+            }),
+            "history": spaces.Text(500),
+            "waiting_room": spaces.Discrete(100),
+            "available_beds": spaces.Discrete(20),
+        })
+        
+        # State
+        self.current_patient: Optional[Patient] = None
+        self.waiting_queue: list = []
+        self.step_count: int = 0
+        self.total_reward: float = 0.0
+        self.available_beds: int = 10
+        self.episode_stats: Dict[str, Any] = {}
+        
+    def reset(
+        self,
+        seed: Optional[int] = None,
+        options: Optional[Dict] = None,
+    ) -> Tuple[Dict, Dict]:
+        """Reset the environment to initial state."""
+        super().reset(seed=seed)
+        
+        if seed is not None:
+            self.patient_generator = PatientGenerator(seed=seed)
+        
+        # Reset state
+        self.step_count = 0
+        self.total_reward = 0.0
+        self.available_beds = 10
+        self.episode_stats = {
+            "correct_triage": 0,
+            "safety_failures": 0,
+            "patients_seen": 0,
+        }
+        
+        # Generate initial waiting room
+        initial_patients = np.random.randint(3, 8)
+        self.waiting_queue = self.patient_generator.generate_batch(initial_patients)
+        for i, p in enumerate(self.waiting_queue):
+            p.time_arrived = -i * 5  # Stagger arrival times
+        
+        # Get first patient
+        self.current_patient = self._get_next_patient()
+        
+        return self._get_observation(), self._get_info()
+    
+    def step(self, action: Dict) -> Tuple[Dict, float, bool, bool, Dict]:
+        """
+        Execute one step in the environment.
+        
+        Args:
+            action: Dict with 'triage_category' (1-5) and 'intervention' (index)
+        
+        Returns:
+            observation, reward, terminated, truncated, info
+        """
+        self.step_count += 1
+        
+        if self.current_patient is None:
+            # No more patients - episode ends
+            return self._get_observation(), 0.0, True, False, self._get_info()
+        
+        # Parse action
+        assigned_category = action.get("triage_category", 3)
+        intervention_idx = action.get("intervention", 0)
+        intervention = self.INTERVENTIONS[intervention_idx]
+        
+        # Calculate reward
+        reward = self._calculate_reward(assigned_category, intervention)
+        self.total_reward += reward
+        self.episode_stats["patients_seen"] += 1
+        
+        # Update bed availability based on intervention
+        if intervention in ["send_to_resus", "send_to_majors"]:
+            self.available_beds = max(0, self.available_beds - 1)
+        elif intervention in ["discharge", "refer_to_gp"]:
+            self.available_beds = min(10, self.available_beds + 1)
+        
+        # Possibly add new patients to queue
+        if np.random.random() < 0.3:  # 30% chance of new arrival
+            new_patient = self.patient_generator.generate()
+            new_patient.time_arrived = self.step_count
+            self.waiting_queue.append(new_patient)
+        
+        # Get next patient
+        self.current_patient = self._get_next_patient()
+        
+        # Check termination
+        terminated = self.current_patient is None and len(self.waiting_queue) == 0
+        truncated = self.step_count >= self.max_steps
+        
+        return self._get_observation(), reward, terminated, truncated, self._get_info()
+    
+    def _calculate_reward(self, assigned_category: int, intervention: str) -> float:
+        """Calculate reward based on triage decision."""
+        if self.current_patient is None:
+            return 0.0
+        
+        true_category = self.current_patient.true_category
+        category_diff = abs(assigned_category - true_category)
+        
+        reward = 0.0
+        
+        # Category accuracy
+        if category_diff == 0:
+            reward += 10.0
+            self.episode_stats["correct_triage"] += 1
+        elif category_diff == 1:
+            reward += 5.0  # Close enough
+        else:
+            reward -= 5.0 * category_diff  # Penalty scales with error
+        
+        # Critical safety failure: Under-triaging a critical patient
+        if true_category <= 2 and assigned_category >= true_category + 2:
+            reward -= 50.0
+            self.episode_stats["safety_failures"] += 1
+        
+        # Intervention appropriateness
+        if true_category == 1 and intervention == "send_to_resus":
+            reward += 5.0
+        elif true_category == 5 and intervention in ["discharge", "refer_to_gp"]:
+            reward += 3.0
+        elif true_category == 1 and intervention == "discharge":
+            reward -= 30.0  # Never discharge a P1!
+        
+        return reward
+    
+    def _get_next_patient(self) -> Optional[Patient]:
+        """Get the next patient from the queue (FIFO with priority override)."""
+        if not self.waiting_queue:
+            return None
+        
+        # Priority override: P1 patients jump the queue
+        for i, patient in enumerate(self.waiting_queue):
+            if patient.true_category == 1:
+                return self.waiting_queue.pop(i)
+        
+        # Otherwise FIFO
+        return self.waiting_queue.pop(0)
+    
+    def _get_observation(self) -> Dict:
+        """Build the observation dictionary."""
+        if self.current_patient is None:
+            return {
+                "patient_id": "",
+                "chief_complaint": "No patients waiting.",
+                "vitals": {
+                    "hr": 0.0, "bp_sys": 0.0, "bp_dia": 0.0,
+                    "spo2": 0.0, "rr": 0.0, "temp": 0.0, "avpu": "A"
+                },
+                "history": "",
+                "waiting_room": len(self.waiting_queue),
+                "available_beds": self.available_beds,
+            }
+        
+        return {
+            "patient_id": self.current_patient.id,
+            "chief_complaint": self.current_patient.chief_complaint,
+            "vitals": {
+                "hr": float(self.current_patient.vitals.get("hr", 0)),
+                "bp_sys": float(self.current_patient.vitals.get("bp_sys", 0)),
+                "bp_dia": float(self.current_patient.vitals.get("bp_dia", 0)),
+                "spo2": float(self.current_patient.vitals.get("spo2", 0)),
+                "rr": float(self.current_patient.vitals.get("rr", 0)),
+                "temp": float(self.current_patient.vitals.get("temp", 0)),
+                "avpu": str(self.current_patient.vitals.get("avpu", "A")),
+            },
+            "history": self.current_patient.history,
+            "waiting_room": len(self.waiting_queue),
+            "available_beds": self.available_beds,
+        }
+    
+    def _get_info(self) -> Dict:
+        """Return additional info."""
+        return {
+            "step": self.step_count,
+            "total_reward": self.total_reward,
+            "true_category": self.current_patient.true_category if self.current_patient else None,
+            **self.episode_stats,
+        }
+    
+    def render(self) -> Optional[str]:
+        """Render the environment."""
+        if self.render_mode == "human" or self.render_mode == "ansi":
+            obs = self._get_observation()
+            output = f"""
+╔══════════════════════════════════════════════════════════════════╗
+║  A&E TRIAGE SIMULATOR  │  Step: {self.step_count:3d} │ Waiting: {obs['waiting_room']:2d} │ Beds: {obs['available_beds']:2d}  ║
+╠══════════════════════════════════════════════════════════════════╣
+║  PATIENT: {obs['patient_id']:<54} ║
+╠──────────────────────────────────────────────────────────────────╣
+║  Chief Complaint:                                                ║
+║    "{obs['chief_complaint'][:60]:<60}" ║
+╠──────────────────────────────────────────────────────────────────╣
+║  VITALS:                                                         ║
+║    HR: {obs['vitals']['hr']:>3.0f}  │  BP: {obs['vitals']['bp_sys']:>3.0f}/{obs['vitals']['bp_dia']:<3.0f}  │  SpO2: {obs['vitals']['spo2']:>3.0f}%            ║
+║    RR: {obs['vitals']['rr']:>3.0f}  │  Temp: {obs['vitals']['temp']:.1f}°C  │  AVPU: {obs['vitals']['avpu']}               ║
+╠──────────────────────────────────────────────────────────────────╣
+║  History: {obs['history'][:55]:<55} ║
+╠══════════════════════════════════════════════════════════════════╣
+║  What is your triage decision?                                   ║
+║    [1] Immediate  [2] Very Urgent  [3] Urgent  [4] Std  [5] Non  ║
+╚══════════════════════════════════════════════════════════════════╝
+"""
+            if self.render_mode == "human":
+                print(output)
+            return output
+        return None
+    
+    def close(self):
+        """Clean up resources."""
+        pass
+
+
+# Register with Gymnasium
+gym.register(
+    id="NurseSim-Triage-v0",
+    entry_point="nursesim_rl:TriageEnv",
+)

package.json

@@ -0,0 +1,18 @@
+{
+    "name": "nursesim-rl",
+    "version": "0.1.0",
+    "description": "A Triage Environment for Reinforcement Learning - OpenEnv Challenge Entry",
+    "author": "Lincoln Gombedza",
+    "license": "MIT",
+    "keywords": [
+        "reinforcement-learning",
+        "nursing",
+        "triage",
+        "openenv",
+        "gymnasium"
+    ],
+    "dependencies": {
+        "gymnasium": ">=0.29.0",
+        "numpy": ">=1.24.0"
+    }
+}

push_dataset.py

@@ -0,0 +1,33 @@
+
+import os
+import sys
+from huggingface_hub import HfApi, create_repo
+from datasets import load_dataset
+
+def push_dataset(file_path, repo_id, token=None):
+    """
+    Pushes a JSONL dataset to Hugging Face Hub
+    """
+    print(f"🚀 Loading dataset from {file_path}...")
+    try:
+        dataset = load_dataset('json', data_files=file_path)
+    except Exception as e:
+        print(f"❌ Error loading dataset: {e}")
+        return
+
+    print(f"📦 Pushing to {repo_id}...")
+    try:
+        dataset.push_to_hub(repo_id, token=token)
+        print(f"✅ Successfully pushed to https://huggingface.co/datasets/{repo_id}")
+    except Exception as e:
+        print(f"❌ Error pushing to Hub: {e}")
+
+if __name__ == "__main__":
+    if len(sys.argv) < 3:
+        print("Usage: python push_dataset.py <file_path> <repo_id>")
+        sys.exit(1)
+    
+    file_path = sys.argv[1]
+    repo_id = sys.argv[2]
+    
+    push_dataset(file_path, repo_id)

requirements.txt

@@ -1,12 +1,12 @@
-# NurseSim-Triage Gradio Demo - Hugging Face Spaces Requirements
-# Compatible with ZeroGPU (No Unsloth - uses standard Transformers+PEFT)
-
-gradio
-spaces
-torch
-transformers
-peft
-bitsandbytes
-accelerate
-agentbeats
-gymnasium
+# NurseSim-Triage Gradio Demo - Hugging Face Spaces Requirements
+# Compatible with ZeroGPU (No Unsloth - uses standard Transformers+PEFT)
+
+gradio
+spaces
+torch
+transformers
+peft
+bitsandbytes
+accelerate
+agentbeats
+gymnasium

test_env.py

@@ -0,0 +1,103 @@
+"""
+Test script: Verify the Triage Environment works correctly
+
+Run: python test_env.py
+"""
+
+import sys
+sys.path.insert(0, '.')
+
+from nursesim_rl import TriageEnv, PatientGenerator
+
+
+def test_patient_generator():
+    """Test the patient generator."""
+    print("Testing PatientGenerator...")
+    gen = PatientGenerator(seed=42)
+    
+    for category in range(1, 6):
+        patient = gen.generate(category=category)
+        assert patient.true_category == category
+        assert len(patient.chief_complaint) > 0
+        assert "hr" in patient.vitals
+        print(f"  [OK] Category {category}: {patient.chief_complaint[:40]}...")
+    
+    print("  [OK] PatientGenerator tests passed!\n")
+
+
+def test_triage_env():
+    """Test the triage environment."""
+    print("Testing TriageEnv...")
+    
+    env = TriageEnv(seed=42)
+    obs, info = env.reset()
+    
+    assert "patient_id" in obs
+    assert "chief_complaint" in obs
+    assert "vitals" in obs
+    assert "waiting_room" in obs
+    print(f"  [OK] Reset works, first patient: {obs['patient_id']}")
+    
+    # Take some steps
+    for i in range(5):
+        action = {
+            "triage_category": 3,  # Default to Urgent
+            "intervention": 1,     # Send to majors
+        }
+        obs, reward, terminated, truncated, info = env.step(action)
+        print(f"  [OK] Step {i+1}: Reward={reward:.1f}, Waiting={obs['waiting_room']}")
+        
+        if terminated or truncated:
+            break
+    
+    env.close()
+    print("  [OK] TriageEnv tests passed!\n")
+
+
+def test_reward_calculation():
+    """Test reward calculations."""
+    print("Testing Reward Logic...")
+    
+    env = TriageEnv(seed=123)
+    obs, info = env.reset()
+    
+    # Force a specific patient for testing
+    from nursesim_rl.patient_generator import Patient
+    test_patient = Patient(
+        id="TEST001",
+        chief_complaint="Test complaint",
+        vitals={"hr": 100, "bp_sys": 120, "bp_dia": 80, "spo2": 98, "rr": 16, "temp": 37.0, "avpu": "A"},
+        history="Test history",
+        true_category=1,  # Critical patient!
+        time_arrived=0,
+    )
+    env.current_patient = test_patient
+    
+    # Test correct triage
+    action = {"triage_category": 1, "intervention": 0}  # Correct: Cat 1, Resus
+    _, reward, _, _, _ = env.step(action)
+    print(f"  Correct triage (Cat 1): Reward = {reward:.1f} (expected +15)")
+    
+    # Reset and test safety failure
+    env.reset()
+    env.current_patient = test_patient
+    action = {"triage_category": 4, "intervention": 5}  # Wrong: Cat 4, Discharge (DANGEROUS!)
+    _, reward, _, _, _ = env.step(action)
+    print(f"  Safety failure (Cat 1 -> 4 + Discharge): Reward = {reward:.1f} (expected negative)")
+    
+    env.close()
+    print("  [OK] Reward logic tests passed!\n")
+
+
+if __name__ == "__main__":
+    print("\n" + "="*60)
+    print("[TEST] NURSESIM-RL TEST SUITE")
+    print("="*60 + "\n")
+    
+    test_patient_generator()
+    test_triage_env()
+    test_reward_calculation()
+    
+    print("="*60)
+    print("[PASS] ALL TESTS PASSED!")
+    print("="*60 + "\n")

test_semantic.py

@@ -0,0 +1,64 @@
+"""
+Test script for the Semantic RL Environment.
+Validates that the NurseEmbedWrapper works correctly.
+"""
+
+import numpy as np
+from nursesim_rl import TriageEnv, NurseEmbedWrapper
+
+def test_semantic_wrapper():
+    print("=" * 60)
+    print("Testing NurseEmbed Semantic RL Wrapper")
+    print("=" * 60)
+    
+    # Create base environment
+    print("\n[1] Creating base TriageEnv...")
+    base_env = TriageEnv(max_steps=10, seed=42)
+    
+    # Wrap with NurseEmbed
+    print("[2] Wrapping with NurseEmbedWrapper...")
+    semantic_env = NurseEmbedWrapper(base_env, use_vitals=True)
+    
+    # Check observation space
+    print(f"\n[3] Observation Space Check:")
+    print(f"    Base Env: {type(base_env.observation_space)}")
+    print(f"    Semantic Env: {semantic_env.observation_space}")
+    print(f"    Expected shape: (390,) [384 embed + 6 vitals]")
+    
+    # Reset and get observation
+    print("\n[4] Resetting environment...")
+    obs, info = semantic_env.reset(seed=42)
+    
+    print(f"    Observation type: {type(obs)}")
+    print(f"    Observation shape: {obs.shape}")
+    print(f"    Observation range: [{obs.min():.3f}, {obs.max():.3f}]")
+    
+    # Take a step
+    print("\n[5] Taking a step with action (Cat=3, Intervention=2)...")
+    action = {"triage_category": 3, "intervention": 2}
+    obs2, reward, terminated, truncated, info = semantic_env.step(action)
+    
+    print(f"    Reward: {reward}")
+    print(f"    Terminated: {terminated}")
+    print(f"    New observation shape: {obs2.shape}")
+    
+    # Verify embedding is meaningful (not just zeros)
+    print("\n[6] Embedding quality check:")
+    embed_part = obs[:384]  # First 384 dims are the embedding
+    print(f"    Embedding L2 norm: {np.linalg.norm(embed_part):.3f}")
+    print(f"    Embedding is normalized: {abs(np.linalg.norm(embed_part) - 1.0) < 0.1}")
+    
+    # Test caching
+    print("\n[7] Testing embedding cache...")
+    obs3, _ = semantic_env.reset(seed=42)  # Same seed = same patient
+    cache_hit = np.allclose(obs[:384], obs3[:384])
+    print(f"    Cache working: {cache_hit}")
+    
+    print("\n" + "=" * 60)
+    print("ALL TESTS PASSED!")
+    print("=" * 60)
+    
+    return True
+
+if __name__ == "__main__":
+    test_semantic_wrapper()

tests/test_a2a_compliance.py

@@ -0,0 +1,184 @@
+#!/usr/bin/env python3
+"""
+A2A Protocol Compliance Tests for NurseSim-Triage Agent
+
+Tests the agent's conformance to the Agent-to-Agent (A2A) protocol specification.
+"""
+
+import json
+import subprocess
+import time
+import requests
+from pathlib import Path
+
+
+def test_agent_card_exists():
+    """Test that agent-card.json exists and is valid JSON."""
+    agent_card_path = Path(".well-known/agent-card.json")
+    assert agent_card_path.exists(), "agent-card.json not found"
+    
+    with open(agent_card_path) as f:
+        card = json.load(f)
+    
+    # Validate required fields
+    assert "name" in card, "Missing 'name' field"
+    assert "protocol" in card, "Missing 'protocol' field"
+    assert card["protocol"] == "a2a/v1.0", f"Invalid protocol version: {card['protocol']}"
+    assert "capabilities" in card, "Missing 'capabilities' field"
+    assert "input_schema" in card["capabilities"], "Missing input_schema"
+    assert "output_schema" in card["capabilities"], "Missing output_schema"
+    
+    print("✓ Agent card validation passed")
+    return card
+
+
+def test_input_schema_validation(agent_card):
+    """Test that input schema is properly defined."""
+    input_schema = agent_card["capabilities"]["input_schema"]
+    
+    assert input_schema["type"] == "object"
+    assert "properties" in input_schema
+    assert "complaint" in input_schema["properties"]
+    assert "vitals" in input_schema["properties"]
+    
+    # Validate vitals schema
+    vitals_schema = input_schema["properties"]["vitals"]
+    assert "heart_rate" in vitals_schema["properties"]
+    assert "blood_pressure" in vitals_schema["properties"]
+    assert "spo2" in vitals_schema["properties"]
+    assert "temperature" in vitals_schema["properties"]
+    
+    print("✓ Input schema validation passed")
+
+
+def test_output_schema_validation(agent_card):
+    """Test that output schema is properly defined."""
+    output_schema = agent_card["capabilities"]["output_schema"]
+    
+    assert output_schema["type"] == "object"
+    assert "properties" in output_schema
+    assert "triage_category" in output_schema["properties"]
+    assert "assessment" in output_schema["properties"]
+    
+    # Validate triage category enum
+    triage_prop = output_schema["properties"]["triage_category"]
+    assert "enum" in triage_prop
+    expected_categories = ["Immediate", "Very Urgent", "Urgent", "Standard", "Non-Urgent"]
+    assert triage_prop["enum"] == expected_categories
+    
+    print("✓ Output schema validation passed")
+
+
+def test_agent_task_processing():
+    """Test that the agent can process a task and return valid output."""
+    # This test requires the agent to be running
+    # For CI/CD, we'll test the agent_main.py module directly
+    
+    try:
+        from agent_main import NurseSimTriageAgent
+        
+        # Initialize agent (requires HF_TOKEN in environment)
+        agent = NurseSimTriageAgent()
+        
+        # Test task
+        test_task = {
+            "complaint": "Severe chest pain radiating to left arm",
+            "vitals": {
+                "heart_rate": 115,
+                "blood_pressure": "85/60",
+                "spo2": 91,
+                "temperature": 37.5
+            }
+        }
+        
+        # Process task
+        result = agent.process_task(test_task)
+        
+        # Validate result structure
+        assert "triage_category" in result, "Missing triage_category in result"
+        assert "assessment" in result, "Missing assessment in result"
+        assert isinstance(result["assessment"], str), "Assessment must be a string"
+        
+        # Validate triage category is valid
+        valid_categories = ["Immediate", "Very Urgent", "Urgent", "Standard", "Non-Urgent", "Error"]
+        assert result["triage_category"] in valid_categories, \
+            f"Invalid triage category: {result['triage_category']}"
+        
+        print(f"✓ Agent task processing passed")
+        print(f"  Triage: {result['triage_category']}")
+        print(f"  Assessment: {result['assessment'][:100]}...")
+        
+        return True
+        
+    except ImportError:
+        print("⚠ Skipping task processing test (agent_main.py not available)")
+        return False
+    except ValueError as e:
+        if "HF_TOKEN" in str(e):
+            print("⚠ Skipping task processing test (HF_TOKEN not set)")
+            return False
+        raise
+
+
+def test_agent_health_check():
+    """Test agent health check endpoint."""
+    try:
+        from agent_main import NurseSimTriageAgent
+        
+        agent = NurseSimTriageAgent()
+        health = agent.health_check()
+        
+        assert "status" in health
+        assert health["status"] == "healthy"
+        assert "model_loaded" in health
+        
+        print("✓ Health check passed")
+        return True
+        
+    except (ImportError, ValueError):
+        print("⚠ Skipping health check test")
+        return False
+
+
+def run_all_tests():
+    """Run all A2A compliance tests."""
+    print("\n" + "="*60)
+    print("NurseSim-Triage A2A Protocol Compliance Tests")
+    print("="*60 + "\n")
+    
+    try:
+        # Test 1: Agent card validation
+        agent_card = test_agent_card_exists()
+        
+        # Test 2: Input schema validation
+        test_input_schema_validation(agent_card)
+        
+        # Test 3: Output schema validation
+        test_output_schema_validation(agent_card)
+        
+        # Test 4: Task processing (requires model)
+        test_agent_task_processing()
+        
+        # Test 5: Health check
+        test_agent_health_check()
+        
+        print("\n" + "="*60)
+        print("✓ All A2A compliance tests passed!")
+        print("="*60 + "\n")
+        
+        return True
+        
+    except AssertionError as e:
+        print(f"\n✗ Test failed: {e}")
+        return False
+    except Exception as e:
+        print(f"\n✗ Unexpected error: {e}")
+        import traceback
+        traceback.print_exc()
+        return False
+
+
+if __name__ == "__main__":
+    import sys
+    success = run_all_tests()
+    sys.exit(0 if success else 1)

tests/test_dual_mode.sh

@@ -0,0 +1,70 @@
+#!/bin/bash
+# Dual-Mode Configuration Test
+# Tests that the agent correctly switches between Gradio and A2A modes
+
+set -e
+
+echo "========================================"
+echo "NurseSim-Triage Dual-Mode Test"
+echo "========================================"
+echo ""
+
+# Test 1: A2A Mode
+echo "Test 1: Setting AGENT_MODE=a2a"
+export AGENT_MODE=a2a
+# Run in background and redirect to file
+./run.sh > a2a_output.log 2>&1 &
+pid=$!
+echo "Process started with PID: $pid"
+sleep 5
+# Kill the process
+kill $pid 2>/dev/null || true
+
+# Check output file
+if grep -q "AgentBeats A2A mode" a2a_output.log; then
+    echo "✓ A2A mode launched correctly"
+    cat a2a_output.log
+    rm a2a_output.log
+else
+    echo "✗ A2A mode failed. Log output:"
+    cat a2a_output.log
+    rm a2a_output.log || true
+    exit 1
+fi
+
+# Test 2: Gradio Mode
+echo ""
+echo "Test 2: Setting AGENT_MODE=gradio"
+export AGENT_MODE=gradio
+./run.sh > gradio_output.log 2>&1 &
+pid=$!
+echo "Process started with PID: $pid"
+sleep 5
+kill $pid 2>/dev/null || true
+
+if grep -q "Gradio demo" gradio_output.log; then
+    echo "✓ Gradio mode launched correctly"
+    cat gradio_output.log
+    rm gradio_output.log
+else
+    echo "✗ Gradio mode failed. Log output:"
+    cat gradio_output.log
+    rm gradio_output.log || true
+    exit 1
+fi
+
+# Test 3: Invalid Mode
+echo ""
+echo "Test 3: Invalid AGENT_MODE"
+export AGENT_MODE=invalid
+if ./run.sh 2>&1 | grep -q "Error: Invalid AGENT_MODE"; then
+    echo "✓ Invalid mode properly rejected"
+else
+    echo "✗ Invalid mode handling failed"
+    exit 1
+fi
+
+echo ""
+echo "========================================"
+echo "✓ All dual-mode tests passed!"
+echo "========================================"

train_semantic_agent.py

@@ -0,0 +1,143 @@
+"""
+Train a PPO Agent on the Semantic Triage Environment.
+
+This script trains an agent that learns from NurseEmbed-encoded clinical observations.
+"""
+
+import os
+import numpy as np
+from datetime import datetime
+
+# Stable Baselines 3
+from stable_baselines3 import PPO
+from stable_baselines3.common.env_util import make_vec_env
+from stable_baselines3.common.callbacks import EvalCallback, BaseCallback
+from stable_baselines3.common.vec_env import DummyVecEnv
+
+# Our environment
+from nursesim_rl import TriageEnv, NurseEmbedWrapper
+
+
+class PrintProgressCallback(BaseCallback):
+    """Simple callback to print training progress."""
+    
+    def __init__(self, print_freq: int = 1000, verbose: int = 0):
+        super().__init__(verbose)
+        self.print_freq = print_freq
+        
+    def _on_step(self) -> bool:
+        if self.n_calls % self.print_freq == 0:
+            # Get recent episode rewards if available
+            if len(self.model.ep_info_buffer) > 0:
+                mean_reward = np.mean([ep['r'] for ep in self.model.ep_info_buffer])
+                mean_length = np.mean([ep['l'] for ep in self.model.ep_info_buffer])
+                print(f"Step {self.n_calls}: Mean Reward = {mean_reward:.2f}, Mean Length = {mean_length:.1f}")
+        return True
+
+
+def make_semantic_env():
+    """Factory function for creating the semantic environment."""
+    base_env = TriageEnv(max_steps=50, max_patients=20)
+    semantic_env = NurseEmbedWrapper(base_env, use_vitals=True)
+    return semantic_env
+
+
+def train_semantic_agent(
+    total_timesteps: int = 50000,
+    save_path: str = "models/semantic_ppo",
+    log_dir: str = "logs/semantic_ppo",
+):
+    """Train a PPO agent on the semantic triage environment."""
+    
+    print("=" * 60)
+    print("SEMANTIC RL TRAINING")
+    print(f"Started: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
+    print("=" * 60)
+    
+    # Create directories
+    os.makedirs(save_path, exist_ok=True)
+    os.makedirs(log_dir, exist_ok=True)
+    
+    # Create vectorized environment
+    print("\n[1] Creating Semantic Environment...")
+    env = DummyVecEnv([make_semantic_env])
+    
+    print(f"    Observation space: {env.observation_space}")
+    print(f"    Action space: {env.action_space}")
+    
+    # Create evaluation environment
+    eval_env = DummyVecEnv([make_semantic_env])
+    
+    # Create the PPO agent
+    print("\n[2] Initializing PPO Agent...")
+    model = PPO(
+        "MlpPolicy",
+        env,
+        learning_rate=3e-4,
+        n_steps=2048,
+        batch_size=64,
+        n_epochs=10,
+        gamma=0.99,
+        gae_lambda=0.95,
+        clip_range=0.2,
+        ent_coef=0.01,
+        verbose=0,
+        tensorboard_log=log_dir,
+    )
+    
+    print(f"    Policy architecture: {model.policy}")
+    
+    # Callbacks
+    progress_callback = PrintProgressCallback(print_freq=2000)
+    eval_callback = EvalCallback(
+        eval_env,
+        best_model_save_path=save_path,
+        log_path=log_dir,
+        eval_freq=5000,
+        n_eval_episodes=5,
+        deterministic=True,
+        verbose=0,
+    )
+    
+    # Train!
+    print(f"\n[3] Training for {total_timesteps:,} timesteps...")
+    print("-" * 60)
+    
+    model.learn(
+        total_timesteps=total_timesteps,
+        callback=[progress_callback, eval_callback],
+        progress_bar=True,
+    )
+    
+    # Save final model
+    final_path = os.path.join(save_path, "semantic_ppo_final")
+    model.save(final_path)
+    print(f"\n[4] Model saved to: {final_path}")
+    
+    # Quick evaluation
+    print("\n[5] Final Evaluation (10 episodes)...")
+    rewards = []
+    for i in range(10):
+        obs = eval_env.reset()
+        episode_reward = 0
+        done = False
+        while not done:
+            action, _ = model.predict(obs, deterministic=True)
+            obs, reward, done, info = eval_env.step(action)
+            episode_reward += reward[0]
+        rewards.append(episode_reward)
+    
+    print(f"    Mean Reward: {np.mean(rewards):.2f} +/- {np.std(rewards):.2f}")
+    print(f"    Best Episode: {np.max(rewards):.2f}")
+    print(f"    Worst Episode: {np.min(rewards):.2f}")
+    
+    print("\n" + "=" * 60)
+    print("TRAINING COMPLETE!")
+    print("=" * 60)
+    
+    return model
+
+
+if __name__ == "__main__":
+    # Run with reduced timesteps for quick demo
+    train_semantic_agent(total_timesteps=20000)

viz_semantic.py

@@ -0,0 +1,173 @@
+"""
+Visualize "What the Agent Sees" - Semantic Embedding Projection
+
+This script generates a 2D t-SNE visualization of how the RL agent
+perceives clinical observations through NurseEmbed semantic vectors.
+"""
+
+import numpy as np
+import matplotlib.pyplot as plt
+from sklearn.manifold import TSNE
+from sklearn.decomposition import PCA
+import os
+
+from nursesim_rl import TriageEnv, NurseEmbedWrapper
+
+
+def collect_observations(n_observations: int = 200, seed: int = 42):
+    """Collect observations and their true categories."""
+    
+    print(f"[1] Collecting {n_observations} observations...")
+    
+    # Create semantic environment
+    base_env = TriageEnv(max_steps=100, max_patients=50, seed=seed)
+    env = NurseEmbedWrapper(base_env, use_vitals=True)
+    
+    observations = []
+    categories = []
+    complaints = []
+    
+    obs, info = env.reset(seed=seed)
+    
+    for i in range(n_observations):
+        # Store observation and metadata
+        observations.append(obs.copy())
+        
+        # Get true category from wrapped env
+        if base_env.current_patient:
+            categories.append(base_env.current_patient.true_category)
+            complaints.append(base_env.current_patient.chief_complaint[:50])
+        else:
+            categories.append(3)  # Default
+            complaints.append("No patient")
+        
+        # Take a random action to move to next patient
+        action = env.action_space.sample()
+        obs, _, terminated, truncated, _ = env.step(action)
+        
+        if terminated or truncated:
+            obs, _ = env.reset()
+    
+    print(f"    Collected {len(observations)} observations")
+    
+    return np.array(observations), np.array(categories), complaints
+
+
+def visualize_embeddings(observations, categories, complaints, output_path="viz"):
+    """Create t-SNE visualization of embeddings."""
+    
+    print("[2] Computing t-SNE projection...")
+    
+    # Extract just the embedding part (first 384 dims)
+    embeddings = observations[:, :384]
+    
+    # First reduce with PCA for speed (384 -> 50)
+    pca = PCA(n_components=50)
+    embeddings_pca = pca.fit_transform(embeddings)
+    print(f"    PCA explained variance: {sum(pca.explained_variance_ratio_):.2%}")
+    
+    # Then t-SNE to 2D
+    tsne = TSNE(n_components=2, perplexity=30, random_state=42, max_iter=1000)
+    embeddings_2d = tsne.fit_transform(embeddings_pca)
+    
+    print("[3] Creating visualization...")
+    
+    # Create figure
+    fig, ax = plt.subplots(figsize=(12, 10))
+    
+    # Color map for triage categories
+    colors = {
+        1: '#FF0000',  # Immediate - Red
+        2: '#FF8C00',  # Very Urgent - Orange
+        3: '#FFD700',  # Urgent - Yellow
+        4: '#32CD32',  # Standard - Green
+        5: '#1E90FF',  # Non-urgent - Blue
+    }
+    
+    category_names = {
+        1: 'P1: Immediate',
+        2: 'P2: Very Urgent',
+        3: 'P3: Urgent',
+        4: 'P4: Standard',
+        5: 'P5: Non-urgent'
+    }
+    
+    # Plot each category
+    for cat in sorted(set(categories)):
+        mask = categories == cat
+        ax.scatter(
+            embeddings_2d[mask, 0],
+            embeddings_2d[mask, 1],
+            c=colors.get(cat, '#888888'),
+            label=category_names.get(cat, f'Category {cat}'),
+            alpha=0.7,
+            s=100,
+            edgecolors='white',
+            linewidths=0.5
+        )
+    
+    # Styling
+    ax.set_title(
+        'What the Agent Sees: Semantic Patient Clusters\n'
+        '(t-SNE projection of NurseEmbed vectors)',
+        fontsize=14, fontweight='bold'
+    )
+    ax.set_xlabel('Semantic Dimension 1', fontsize=11)
+    ax.set_ylabel('Semantic Dimension 2', fontsize=11)
+    
+    # Legend
+    ax.legend(loc='upper right', title='Triage Category', fontsize=10)
+    
+    # Clean up axes
+    ax.spines['top'].set_visible(False)
+    ax.spines['right'].set_visible(False)
+    ax.grid(True, alpha=0.3)
+    
+    # Add annotation
+    ax.text(
+        0.02, 0.02,
+        'Patients with similar clinical presentations\ncluster together in semantic space.',
+        transform=ax.transAxes,
+        fontsize=9,
+        verticalalignment='bottom',
+        style='italic',
+        bbox=dict(boxstyle='round', facecolor='wheat', alpha=0.5)
+    )
+    
+    # Save
+    os.makedirs(output_path, exist_ok=True)
+    output_file = os.path.join(output_path, 'semantic_clusters.png')
+    plt.tight_layout()
+    plt.savefig(output_file, dpi=150, bbox_inches='tight')
+    print(f"[4] Saved to: {output_file}")
+    
+    # Also show statistics
+    print("\n[5] Cluster Statistics:")
+    for cat in sorted(set(categories)):
+        mask = categories == cat
+        center = embeddings_2d[mask].mean(axis=0)
+        print(f"    {category_names[cat]}: {mask.sum()} patients, center at ({center[0]:.1f}, {center[1]:.1f})")
+    
+    return output_file
+
+
+def main():
+    print("=" * 60)
+    print("SEMANTIC EMBEDDING VISUALIZATION")
+    print("'What the Agent Sees'")
+    print("=" * 60 + "\n")
+    
+    # Collect data
+    observations, categories, complaints = collect_observations(n_observations=150)
+    
+    # Visualize
+    output_file = visualize_embeddings(observations, categories, complaints)
+    
+    print("\n" + "=" * 60)
+    print("VISUALIZATION COMPLETE!")
+    print(f"Open: {os.path.abspath(output_file)}")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()