Update app.py

app.py

@@ -1,25 +1,76 @@
-from fastapi import FastAPI
-from pydantic import BaseModel
-from transformers import T5Tokenizer, T5ForConditionalGeneration
-import torch
+import json
+import random
 
-app = FastAPI()
+# Predefined attributes for empathy training
+emotions = ["lonely", "confused", "sadness", "frustrated", "scared"]
+intents = ["seek_reassurance", "express_confusion", "seek_support", "vent", "request_help"]
+difficulties = ["easy", "moderate", "hard"]
+tags_pool = [
+    ["memory", "forgetting", "support"],
+    ["family", "burden", "dementia"],
+    ["confusion", "orientation", "safety"],
+    ["fear", "loneliness", "aging"],
+    ["support", "validation", "love"]
+]
 
-model = T5ForConditionalGeneration.from_pretrained("obx0x3/empathy-dementia")
-tokenizer = T5Tokenizer.from_pretrained("obx0x3/empathy-dementia")
+# English and French empathetic templates
+en_templates = [
+    ("I feel like {emotion}.", "You're not alone. You're understood."),
+    ("I'm always {emotion}, it's scary.", "That sounds tough. I'm here to support you."),
+    ("Sometimes I just feel so {emotion}.", "You're valued. Let's work through this together."),
+    ("Why do I keep forgetting things?", "It's okay. Memory issues happen, and you’re not alone."),
+    ("I don’t know where I am.", "Let’s take a deep breath. I can help guide you.")
+]
 
-class PromptRequest(BaseModel):
-    message: str
-    lang: str = "en"
+fr_templates = [
+    ("Je me sens {emotion}.", "Tu n'es pas seul. Je suis là pour toi."),
+    ("Je suis toujours {emotion}, c'est effrayant.", "C'est dur. Je te comprends."),
+    ("Parfois je me sens si {emotion}.", "Tu comptes beaucoup. Ensemble, on va y arriver."),
+    ("Pourquoi j'oublie tout le temps ?", "Ce n'est pas grave. Ça arrive à beaucoup de gens."),
+    ("Je ne sais pas où je suis.", "Respire un peu. Je vais t'aider.")
+]
 
-@app.post("/generate")
-async def generate_response(payload: PromptRequest):
-    prefix = "émotion: " if payload.lang == "fr" else ""
-    text = prefix + payload.message
+# Generate a single sample
+def create_entry(lang):
+    templates = en_templates if lang == "en" else fr_templates
+    input_template, response = random.choice(templates)
+    emotion = random.choice(emotions)
+    intent = random.choice(intents)
+    tags = random.choice(tags_pool)
+    difficulty = random.choice(difficulties)
+    input_text = input_template.format(emotion=emotion)
 
-    input_ids = tokenizer.encode(text, return_tensors="pt")
-    with torch.no_grad():
-        outputs = model.generate(input_ids, max_length=50)
+    return {
+        "input": input_text,
+        "response": response,
+        "emotion": emotion,
+        "intent": intent,
+        "tags": tags,
+        "care_mode": True,
+        "language": lang,
+        "difficulty": difficulty,
+        "is_dementia_related": True
+    }
 
-    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
-    return {"reply": result.strip()}
+# Generate 100 English and 70 French samples
+en_data = [create_entry("en") for _ in range(100)]
+fr_data = [create_entry("fr") for _ in range(70)]
+all_data = en_data + fr_data
+random.shuffle(all_data)
+
+# Smart split (70% train, 20% validation, 10% test)
+train = all_data[:int(0.7 * len(all_data))]
+validation = all_data[int(0.7 * len(all_data)):int(0.9 * len(all_data))]
+test = all_data[int(0.9 * len(all_data)):]
+
+# Write to JSON
+with open("dementia_train_split.json", "w", encoding="utf-8") as f:
+    json.dump(train, f, indent=2, ensure_ascii=False)
+
+with open("dementia_validation_split.json", "w", encoding="utf-8") as f:
+    json.dump(validation, f, indent=2, ensure_ascii=False)
+
+with open("dementia_test_multilang.json", "w", encoding="utf-8") as f:
+    json.dump(test, f, indent=2, ensure_ascii=False)
+
+print("✅ Dataset splits created (train/validation/test)")