Upload app.py with huggingface_hub

app.py

@@ -1,6 +1,7 @@
 import gradio as gr
 import torch
 import torch.nn.functional as F
+import re
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 
 MODEL_ID = "gorkem371/pii-intent-classifier-xlmr-large"
@@ -11,37 +12,71 @@ model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
 model.eval()
 print("Model loaded!")
 
-ENTITY_TYPES = [
-    "PHONE", "EMAIL", "SOCIAL_MEDIA", "IBAN", "CREDIT_CARD",
-    "ADDRESS", "URL", "CRYPTO_ADDRESS", "OFF_PLATFORM_ATTEMPT"
-]
+# ── Auto-detect entities from text ──────────────────────────────
 
-EXAMPLES = [
-    ["my number is 05321234567 call me anytime", "05321234567", "PHONE"],
-    ["your order number is ORD-784321", "ORD-784321", "PHONE"],
-    ["numaram 05321234567 bana yazabilirsin", "05321234567", "PHONE"],
-    ["رقمي 0501234567 تواصل معي", "0501234567", "PHONE"],
-    ["follow me on instagram @creator_official", "@creator_official", "SOCIAL_MEDIA"],
-    ["i will send you my whatsapp tomorrow", "NONE", "PHONE"],
-    ["05321234567 this number is a scammer do not call", "05321234567", "PHONE"],
-    ["bake at 180 degrees celsius for 45 minutes", "180", "PHONE"],
-    ["check my bio my number is right there", "NONE", "PHONE"],
-    ["email me at creator@gmail.com", "creator@gmail.com", "EMAIL"],
-    ["my IBAN is GB29NWBK60161331926819", "GB29NWBK60161331926819", "IBAN"],
-    ["siparis numaraniz SPN-78432156", "SPN-78432156", "PHONE"],
-    ["if we agree on terms i will share my number", "NONE", "PHONE"],
-    ["05321234567 times 2 equals what", "05321234567", "PHONE"],
-    ["رقمي 0501234567 هههه كأني بعطيك ياخي بس هو كذا", "0501234567", "PHONE"],
+ENTITY_PATTERNS = [
+    ("EMAIL", r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}'),
+    ("IBAN", r'\b[A-Z]{2}\d{2}[A-Z0-9]{10,30}\b'),
+    ("PHONE", r'(?:\+?\d{1,3}[\s-]?)?\(?\d{2,4}\)?[\s.-]?\d{3,4}[\s.-]?\d{3,4}\b'),
+    ("SOCIAL_MEDIA", r'@[a-zA-Z0-9_.]{2,30}'),
+    ("URL", r'(?:https?://)?(?:www\.)?[a-zA-Z0-9-]+\.[a-zA-Z]{2,}(?:/\S*)?'),
+    ("CRYPTO_ADDRESS", r'\b(?:0x[a-fA-F0-9]{40}|[13][a-km-zA-HJ-NP-Z1-9]{25,34}|bc1[a-zA-HJ-NP-Z0-9]{39,59})\b'),
 ]
 
-
-def classify(context: str, entity: str, entity_type: str):
-    if not context.strip():
-        return {}, "Please enter a message."
-
-    if not entity.strip():
-        entity = "NONE"
-
+INTENT_KEYWORDS = {
+    "PHONE": [
+        "numaram", "numara", "whatsapp", "telefon", "ara beni", "bana ulaş",
+        "رقمي", "واتساب", "اتصل", "كلمني",
+        "my number", "call me", "whatsapp me", "reach me", "contact me",
+        "send you my number", "numara atarim", "numara veririm",
+        "بارسل لك رقمي", "بعطيك رقمي",
+    ],
+    "SOCIAL_MEDIA": [
+        "instagram", "insta", "tiktok", "telegram", "snapchat", "snap",
+        "twitter", "takip et", "follow me", "تابعني", "انستقرام", "سناب",
+        "beni bul", "find me", "تلقاني", "add me",
+    ],
+    "EMAIL": ["mail", "email", "ايميل", "راسلني", "e-posta"],
+    "OFF_PLATFORM_ATTEMPT": [
+        "başka yerde", "platform dışı", "dışarıda", "oradan konuşalım",
+        "let's talk on", "move to", "نروح", "بره", "خارج",
+        "bio", "story", "profil", "بايو", "ستوري",
+    ],
+}
+
+
+def detect_entities(text):
+    """Auto-detect entities in text."""
+    found = []
+
+    for etype, pattern in ENTITY_PATTERNS:
+        for match in re.finditer(pattern, text):
+            found.append((match.group(), etype, match.start()))
+
+    # Remove overlapping matches (keep longest)
+    found.sort(key=lambda x: (-len(x[0]), x[2]))
+    used_ranges = []
+    filtered = []
+    for entity, etype, start in found:
+        end = start + len(entity)
+        if not any(s <= start < e or s < end <= e for s, e in used_ranges):
+            filtered.append((entity, etype))
+            used_ranges.append((start, end))
+
+    return filtered
+
+
+def detect_intent_type(text):
+    """If no entity found, detect intent type from keywords."""
+    text_lower = text.lower()
+    for etype, keywords in INTENT_KEYWORDS.items():
+        if any(kw in text_lower for kw in keywords):
+            return etype
+    return "PHONE"
+
+
+def predict_single(context, entity, entity_type):
+    """Run model on a single entity."""
     text = f"{context} </s> {entity} | {entity_type}"
     inputs = tokenizer(text, max_length=256, padding="max_length", truncation=True, return_tensors="pt")
 
@@ -51,20 +86,79 @@ def classify(context: str, entity: str, entity_type: str):
 
     pii_prob = probs[0][1].item()
     not_pii_prob = probs[0][0].item()
+    return pii_prob, not_pii_prob
+
+
+def classify(message: str):
+    if not message.strip():
+        return {}, "Please enter a message."
+
+    entities = detect_entities(message)
+
+    if not entities:
+        # No explicit entity — check for intent
+        intent_type = detect_intent_type(message)
+        entities = [("NONE", intent_type)]
+
+    # Run model on all detected entities, take the highest PII score
+    best_pii = 0
+    best_not_pii = 1
+    best_entity = ""
+    best_type = ""
+    all_results = []
+
+    for entity, etype in entities:
+        pii_prob, not_pii_prob = predict_single(message, entity, etype)
+        all_results.append((entity, etype, pii_prob))
+        if pii_prob > best_pii:
+            best_pii = pii_prob
+            best_not_pii = not_pii_prob
+            best_entity = entity
+            best_type = etype
 
     label_scores = {
-        "🚨 PII DETECTED": pii_prob,
-        "✅ NOT PII": not_pii_prob,
+        "🚨 PII DETECTED": best_pii,
+        "✅ NOT PII": best_not_pii,
     }
 
-    if pii_prob > not_pii_prob:
-        verdict = f"🚨 **PII DETECTED** — This message contains PII sharing intent.\n\nConfidence: **{pii_prob:.1%}**"
+    # Build verdict
+    if best_pii > best_not_pii:
+        verdict = f"🚨 **PII DETECTED** — This message contains PII sharing intent.\n\n"
+    else:
+        verdict = f"✅ **NOT PII** — No PII sharing intent detected.\n\n"
+
+    verdict += f"**Confidence:** {max(best_pii, best_not_pii):.1%}\n\n"
+
+    # Show what was detected
+    if len(all_results) == 1 and all_results[0][0] == "NONE":
+        verdict += f"📋 No explicit entity found — analyzed message intent as `{best_type}`"
     else:
-        verdict = f"✅ **NOT PII** — No PII sharing intent detected.\n\nConfidence: **{not_pii_prob:.1%}**"
+        verdict += "📋 **Detected entities:**\n"
+        for ent, etype, prob in all_results:
+            icon = "🚨" if prob > 0.5 else "✅"
+            verdict += f"- {icon} `{ent}` ({etype}) → {'PII' if prob > 0.5 else 'NOT PII'} ({prob:.0%})\n"
 
     return label_scores, verdict
 
 
+EXAMPLES = [
+    ["my number is 05321234567 call me anytime"],
+    ["numaram 05321234567 bana yazabilirsin"],
+    ["رقمي 0501234567 تواصل معي"],
+    ["follow me on instagram @creator_official"],
+    ["email me at creator@gmail.com"],
+    ["i will send you my whatsapp tomorrow"],
+    ["check my bio my number is right there"],
+    ["your order number is ORD-784321"],
+    ["05321234567 this number is a scammer do not call"],
+    ["bake at 180 degrees celsius for 45 minutes"],
+    ["siparis numaraniz SPN-78432156"],
+    ["my number is 00000000000 lol just kidding"],
+    ["رقم الطلب ORD-784321"],
+    ["if we agree on terms i will share my number"],
+    ["my old number was 05321234567 it might still work"],
+]
+
 with gr.Blocks(
     title="PII Intent Classifier",
     theme=gr.themes.Soft(),
@@ -74,64 +168,55 @@ with gr.Blocks(
         # 🔍 PII Intent Classifier
         ### Multilingual context-aware PII detection (Turkish, Arabic, English)
 
-        This model detects whether a message contains **intent to share personal contact information**.
-        Unlike regex-based detection, it understands context: sarcasm, negation, scam warnings, and evasion techniques.
+        Type any message below. The model will automatically detect entities (phone numbers, emails, handles, etc.)
+        and determine if the message contains **intent to share personal contact information**.
 
-        **Model:** [gorkem371/pii-intent-classifier-xlmr-large](https://huggingface.co/gorkem371/pii-intent-classifier-xlmr-large) (XLM-RoBERTa Large, 550M params)
+        **Model:** [gorkem371/pii-intent-classifier-xlmr-large](https://huggingface.co/gorkem371/pii-intent-classifier-xlmr-large) — XLM-RoBERTa Large, 550M params, 97.7% accuracy
         """
     )
 
     with gr.Row():
         with gr.Column(scale=2):
-            context_input = gr.Textbox(
+            message_input = gr.Textbox(
                 label="Message",
-                placeholder="Enter a message to analyze...",
+                placeholder="Type a message to analyze... (Turkish, Arabic, or English)",
                 lines=3,
             )
-            with gr.Row():
-                entity_input = gr.Textbox(
-                    label="Entity",
-                    placeholder="e.g. 05321234567 or NONE",
-                    value="NONE",
-                )
-                entity_type_input = gr.Dropdown(
-                    choices=ENTITY_TYPES,
-                    label="Entity Type",
-                    value="PHONE",
-                )
-            submit_btn = gr.Button("Analyze", variant="primary", size="lg")
+            submit_btn = gr.Button("🔍 Analyze", variant="primary", size="lg")
 
         with gr.Column(scale=1):
             label_output = gr.Label(label="Classification", num_top_classes=2)
-            verdict_output = gr.Markdown(label="Verdict")
+            verdict_output = gr.Markdown(label="Details")
 
     submit_btn.click(
         fn=classify,
-        inputs=[context_input, entity_input, entity_type_input],
+        inputs=[message_input],
         outputs=[label_output, verdict_output],
     )
 
-    context_input.submit(
+    message_input.submit(
         fn=classify,
-        inputs=[context_input, entity_input, entity_type_input],
+        inputs=[message_input],
         outputs=[label_output, verdict_output],
     )
 
     gr.Examples(
         examples=EXAMPLES,
-        inputs=[context_input, entity_input, entity_type_input],
+        inputs=[message_input],
         outputs=[label_output, verdict_output],
         fn=classify,
         cache_examples=False,
-        label="Try these examples",
+        label="💡 Try these examples",
     )
 
     gr.Markdown(
         """
         ---
-        **How it works:** Input format is `{message} </s> {entity} | {entity_type}`. Use `NONE` as entity for implicit PII intent (e.g., "I'll send my number later").
+        **How it works:** The model automatically detects entities (phone numbers, emails, @handles, URLs, IBANs) in your message
+        using regex, then classifies each one for PII sharing intent using a fine-tuned XLM-RoBERTa Large model.
+        If no entity is found, it analyzes the message for implicit intent (e.g., "I'll send my number later").
 
-        **Author:** [Gorkem Yildiz](https://gorkemyildiz.com) | **Accuracy:** 97.7% on 177 multilingual stress tests
+        **Author:** [Gorkem Yildiz](https://gorkemyildiz.com) · **Dataset:** [pii-intent-detection-multilingual](https://huggingface.co/datasets/gorkem371/pii-intent-detection-multilingual)
         """
     )