Update app.py

app.py

@@ -8,54 +8,67 @@ MODEL = "roberta-base-openai-detector"
 tokenizer = AutoTokenizer.from_pretrained(MODEL)
 model = AutoModelForSequenceClassification.from_pretrained(MODEL)
 
+def split_sentences(paragraph):
+    """Split a paragraph into sentences."""
+    return re.split(r'(?<=[.!?]) +', paragraph.strip())
+
 def detect_ai(text):
-    # Split into paragraphs (double newlines)
+    # Split into paragraphs
     paragraphs = re.split(r"\n\s*\n", text.strip())
 
     results = []
+    all_scores = []
+
+    highlighted = ""
     for para in paragraphs:
         if not para.strip():
             continue
-        inputs = tokenizer(para, return_tensors="pt", truncation=True, max_length=512)
-        with torch.no_grad():
-            outputs = model(**inputs)
-            probs = torch.softmax(outputs.logits, dim=1)
 
-        ai_score = float(probs[0][1])  # AI likelihood
-        human_score = 1 - ai_score     # Human likelihood
+        sentences = split_sentences(para)
+        highlighted_para = ""
 
-        # Decide label: if <90% human → AI
-        if human_score < 0.9:
-            label = "🔴 AI"
-            color = "rgb(255,120,120)"  # red
-        else:
-            label = "🟢 Human"
-            color = "rgb(120,255,120)"  # green
+        for sent in sentences:
+            if not sent.strip():
+                continue
 
-        results.append({"paragraph": para, "ai_score": ai_score, "human_score": human_score, "label": label, "color": color})
+            inputs = tokenizer(sent, return_tensors="pt", truncation=True, max_length=512)
+            with torch.no_grad():
+                outputs = model(**inputs)
+                probs = torch.softmax(outputs.logits, dim=1)
 
-    # Build highlighted HTML
-    highlighted = ""
-    for r in results:
-        highlighted += (
-            f"<div style='background-color:{r['color']}; padding:8px; margin-bottom:6px; border-radius:6px'>"
-            f"<b>{r['label']} — Human {round(r['human_score']*100,1)}% | AI {round(r['ai_score']*100,1)}%</b><br>"
-            f"{r['paragraph']}</div>"
-        )
+            ai_score = float(probs[0][1])   # AI likelihood
+            human_score = 1 - ai_score      # Human likelihood
+            all_scores.append(human_score)
+
+            # Decide label
+            if human_score < 0.9:
+                label = "🔴 AI"
+                color = "rgb(255,120,120)"  # red
+            else:
+                label = "🟢 Human"
+                color = "rgb(120,255,120)"  # green
+
+            highlighted_para += (
+                f"<span style='background-color:{color}; padding:2px; border-radius:3px' "
+                f"title='{label} — Human {round(human_score*100,1)}% | AI {round(ai_score*100,1)}%'>"
+                f"{sent} </span>"
+            )
+
+        highlighted += f"<div style='margin-bottom:10px'>{highlighted_para}</div>"
 
     # Compute overall human %
-    if results:
-        avg_human = sum(r['human_score'] for r in results) / len(results)
+    if all_scores:
+        avg_human = sum(all_scores) / len(all_scores)
         total_human = round(avg_human * 100, 2)
         highlighted += f"<p><b>⚖️ Overall Human Probability: {total_human}%</b></p>"
     else:
         total_human = 0.0
 
-    return highlighted, {"paragraphs": results, "overall_human_percent": total_human}
+    return highlighted, {"overall_human_percent": total_human}
 
 with gr.Blocks() as demo:
-    gr.Markdown("## 🤖 AI Detector (Paragraph-level)")
-    gr.Markdown("Paragraphs with <90% human probability are flagged as **AI**.")
+    gr.Markdown("## 🤖 AI Detector (Sentence-level)")
+    gr.Markdown("Each sentence is checked. If Human <90%, it’s flagged as AI.")
     input_text = gr.Textbox(lines=12, placeholder="Paste your essay or report here...")
     output_html = gr.HTML()
     output_json = gr.JSON()