Create app.py

app.py

@@ -0,0 +1,89 @@
+import gradio as gr
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+import torch.nn.functional as F
+import re
+import string
+
+# 1. Model Configuration
+# Updated to point to the final production model
+MODEL_REPO = "SuperSl6/saudi-eou-model-final"
+
+print(f"Loading Model from {MODEL_REPO}...")
+try:
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO)
+    model = AutoModelForSequenceClassification.from_pretrained(MODEL_REPO)
+    model.eval()
+    print("Model loaded successfully.")
+except Exception as e:
+    print(f"Error loading model: {e}")
+
+# 2. Text Normalization Function
+# Matches the preprocessing steps used during training (removing diacritics, normalizing characters)
+def normalize_text(text):
+    text = str(text)
+    # Remove Arabic Diacritics
+    text = re.sub(r'[\u0617-\u061A\u064B-\u0652]', '', text)
+    # Normalize Alef forms to bare Alef
+    text = re.sub(r'[أإآ]', 'ا', text)
+    # Normalize Ya forms
+    text = re.sub(r'ى', 'ي', text)
+    # Remove Punctuation
+    translator = str.maketrans('', '', string.punctuation + '،؛؟')
+    return text.translate(translator).strip()
+
+# 3. Prediction Logic
+def predict_eou(text):
+    if not text or not text.strip():
+        return "Please enter text...", "0%"
+        
+    clean_text = normalize_text(text)
+    
+    # Safety Rules
+    # If the input is less than 2 words, default to WAIT unless it's a specific keyword.
+    if len(clean_text.split()) < 2:
+        whitelist = ["نعم", "لا", "طيب", "سم", "ابشر", "تم", "صحيح", "اكيد"]
+        if clean_text not in whitelist:
+            return "WAIT (Turn Incomplete)", "Safety Rule Triggered"
+
+    # Prepare input for model
+    inputs = tokenizer(clean_text, return_tensors="pt", truncation=True, padding=True)
+    
+    with torch.no_grad():
+        logits = model(**inputs).logits
+        probs = F.softmax(logits, dim=1).numpy()[0]
+    
+    # Label 1 corresponds to COMPLETE (End of Utterance)
+    score_complete = probs[1]
+    
+    # Threshold set to 0.5 as the model was trained on balanced data
+    threshold = 0.5
+    
+    if score_complete >= threshold:
+        decision = "REPLY (Turn Complete)"
+    else:
+        decision = "WAIT (Turn Incomplete)"
+        
+    return decision, f"{score_complete:.1%}"
+
+# 4. Gradio Interface
+examples = [
+    ["السلام عليكم"], 
+    ["ياخي ودي أحجز عند دكتور"], 
+    ["ياخي ودي أحجز موعد عندكم بكرة"]
+]
+
+iface = gr.Interface(
+    fn=predict_eou,
+    inputs=gr.Textbox(label="User Speech Input", placeholder="Type here..."),
+    outputs=[
+        gr.Textbox(label="Agent Decision"),
+        gr.Label(label="Confidence Score")
+    ],
+    title="Saudi Dialect End-of-Utterance (EOU) Detector",
+    description="Final Production Model. Uses SaudiBERT to detect turn-taking in real-time conversation.",
+    examples=examples,
+    allow_flagging="never"
+)
+
+iface.launch()