Update to V11: 99.3% F1, 97% conversation accuracy, new examples

app.py

@@ -142,21 +142,32 @@ def classify(message: str):
 
 
 EXAMPLES = [
+    # PII - Direct sharing
     ["my number is 05321234567 call me anytime"],
     ["numaram 05321234567 bana yazabilirsin"],
     ["رقمي 0501234567 تواصل معي"],
+    # PII - Social media / email
     ["follow me on instagram @creator_official"],
     ["email me at creator@gmail.com"],
+    # PII - Implicit intent
     ["i will send you my whatsapp tomorrow"],
     ["check my bio my number is right there"],
+    ["if we agree on terms i will share my number"],
+    # PII - Humor + real number (V11 improvement)
+    ["numaram pizza siparişi gibi 05321234567 haha 😂"],
+    # PII - Old number still visible
+    ["my old number was 05321234567 it might still work"],
+    # NOT PII - Order/tracking numbers
     ["your order number is ORD-784321"],
+    ["siparis numaraniz SPN-78432156"],
+    # NOT PII - Room number / postal code (V11 improvement)
+    ["oda numaram 532 otelde buluşalım"],
+    ["posta kodu 34720 Kadıköy İstanbul"],
+    # NOT PII - Scam warning / non-contact
     ["05321234567 this number is a scammer do not call"],
     ["bake at 180 degrees celsius for 45 minutes"],
-    ["siparis numaraniz SPN-78432156"],
+    # NOT PII - Sarcasm
     ["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(
@@ -165,13 +176,13 @@ with gr.Blocks(
 ) as demo:
     gr.Markdown(
         """
-        # 🔍 PII Intent Classifier
+        # 🔍 PII Intent Classifier (V11)
         ### Multilingual context-aware PII detection (Turkish, Arabic, English)
 
         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, 97.7% accuracy
+        **Model:** [gorkem371/pii-intent-classifier-xlmr-large](https://huggingface.co/gorkem371/pii-intent-classifier-xlmr-large) — XLM-RoBERTa Large, 550M params | **F1: 99.3%** · **Conversation accuracy: 97.0%** · Trained on 41K samples
         """
     )
 
@@ -216,7 +227,9 @@ with gr.Blocks(
         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) · **Dataset:** [pii-intent-detection-multilingual](https://huggingface.co/datasets/gorkem371/pii-intent-detection-multilingual)
+        **V11 improvements:** Understands humor + real numbers as PII, distinguishes room/postal/tracking numbers from real phone numbers, handles scam warnings with visible numbers, and 6 new NOT-PII categories.
+
+        **Author:** [Gorkem Yildiz](https://gorkemyildiz.com) · **Dataset:** [pii-intent-detection-multilingual](https://huggingface.co/datasets/gorkem371/pii-intent-detection-multilingual) (41,427 samples)
         """
     )