Feat: Use merged MedGemma model (person-centred fine-tuned)

README.md

@@ -17,9 +17,9 @@ An AI-powered tutor designed to guide nursing professionals through the **A-EQUI
 Automatically synced from GitHub via GitHub Actions.
 
 ## 🧠 Model Strategy
-- **Base Model**: `google/gemma-2-2b-it` (fast, ZeroGPU optimized, 2.5B params)
+- **Base Model**: `NurseCitizenDeveloper/relational-intelligence-medgemma-merged` (person-centred, fine-tuned on FONS principles)
 - **Knowledge Base**: RAG implementation using the official PNA A-EQUIP guide.
-- **Persona**: Strong PNA Tutor system prompting for person-centred, restorative voice.
+- **Persona**: Strong PNA Tutor system prompting for restorative supervision focus.
 
 ## ⚖️ Disclaimer
 This tool is for educational and supportive purposes for Professional Nurse Advocates and nursing staff. It does not provide direct clinical advice.

merge_lora_adapter.ipynb

@@ -0,0 +1,120 @@
+{
+    "nbformat": 4,
+    "nbformat_minor": 0,
+    "metadata": {
+        "colab": {
+            "provenance": [],
+            "gpuType": "T4"
+        },
+        "kernelspec": {
+            "name": "python3",
+            "display_name": "Python 3"
+        },
+        "accelerator": "GPU"
+    },
+    "cells": [
+        {
+            "cell_type": "markdown",
+            "source": [
+                "# 🔧 Merge LoRA Adapter to Full Model\n",
+                "\n",
+                "This notebook merges your Unsloth LoRA adapter with its base model and uploads the result to Hugging Face.\n",
+                "\n",
+                "**Before running:**\n",
+                "1. Go to Runtime > Change runtime type > Select **T4 GPU**\n",
+                "2. Have your Hugging Face **Write Token** ready"
+            ],
+            "metadata": {}
+        },
+        {
+            "cell_type": "code",
+            "source": [
+                "# Step 1: Install dependencies\n",
+                "!pip install unsloth transformers accelerate bitsandbytes -q"
+            ],
+            "metadata": {},
+            "execution_count": null,
+            "outputs": []
+        },
+        {
+            "cell_type": "code",
+            "source": [
+                "# Step 2: Login to Hugging Face\n",
+                "from huggingface_hub import login\n",
+                "\n",
+                "# Paste your HF token when prompted\n",
+                "login()"
+            ],
+            "metadata": {},
+            "execution_count": null,
+            "outputs": []
+        },
+        {
+            "cell_type": "code",
+            "source": [
+                "# Step 3: Configuration\n",
+                "# =====================\n",
+                "# Change these values to match your model\n",
+                "\n",
+                "ADAPTER_MODEL = \"NurseCitizenDeveloper/relational-intelligence-unsloth-medgemma\"  # Your LoRA adapter\n",
+                "OUTPUT_NAME = \"NurseCitizenDeveloper/relational-intelligence-medgemma-merged\"    # Where to save merged model\n",
+                "MAX_SEQ_LENGTH = 2048"
+            ],
+            "metadata": {},
+            "execution_count": null,
+            "outputs": []
+        },
+        {
+            "cell_type": "code",
+            "source": [
+                "# Step 4: Load the adapter model\n",
+                "from unsloth import FastLanguageModel\n",
+                "\n",
+                "print(f\"Loading adapter: {ADAPTER_MODEL}\")\n",
+                "\n",
+                "model, tokenizer = FastLanguageModel.from_pretrained(\n",
+                "    model_name=ADAPTER_MODEL,\n",
+                "    max_seq_length=MAX_SEQ_LENGTH,\n",
+                "    dtype=None,  # Auto-detect\n",
+                "    load_in_4bit=True,\n",
+                ")\n",
+                "\n",
+                "print(\"✅ Model loaded successfully!\")"
+            ],
+            "metadata": {},
+            "execution_count": null,
+            "outputs": []
+        },
+        {
+            "cell_type": "code",
+            "source": [
+                "# Step 5: Merge and push to Hugging Face\n",
+                "print(f\"Merging adapter and pushing to: {OUTPUT_NAME}\")\n",
+                "\n",
+                "model.push_to_hub_merged(\n",
+                "    OUTPUT_NAME,\n",
+                "    tokenizer,\n",
+                "    save_method=\"merged_16bit\",  # Full precision merged model\n",
+                ")\n",
+                "\n",
+                "print(\"🎉 Done! Your merged model is now available at:\")\n",
+                "print(f\"https://huggingface.co/{OUTPUT_NAME}\")"
+            ],
+            "metadata": {},
+            "execution_count": null,
+            "outputs": []
+        },
+        {
+            "cell_type": "markdown",
+            "source": [
+                "## ✅ Next Steps\n",
+                "\n",
+                "Once the upload completes:\n",
+                "1. Go to your new model page on Hugging Face\n",
+                "2. Verify it has a `model.safetensors` file\n",
+                "3. Update your PNA Assistant to use the new merged model ID"
+            ],
+            "metadata": {}
+        }
+    ]
+}

pna_client.py

@@ -4,9 +4,8 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 import spaces
 
 class PNAAssistantClient:
-    # Using google/gemma-2-2b-it - a fast, reliable model that works on ZeroGPU
-    # Person-centred voice is achieved through strong PNA system prompting
-    def __init__(self, model_id="google/gemma-2-2b-it"):
+    # Using user's merged MedGemma model - trained on person-centred language
+    def __init__(self, model_id="NurseCitizenDeveloper/relational-intelligence-medgemma-merged"):
         self.model_id = model_id
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
         self.tokenizer = None