Update app.py

app.py

@@ -9,7 +9,6 @@ import gradio as gr
 # MODEL INITIALIZATION
 # -----------------------------
 MODEL_NAME = "fakespot-ai/roberta-base-ai-text-detection-v1"
-
 tokenizer = None
 model = None
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@@ -17,21 +16,18 @@ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 def get_model():
     global tokenizer, model
     if model is None:
-        # Loading messages to help debug in HF logs
         print(f"Loading model: {MODEL_NAME} on {device}")
         tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
-        
-        # Use float32 on CPU to prevent build-time precision hangs
         dtype = torch.float32
         if device.type == "cuda" and torch.cuda.is_bf16_supported():
             dtype = torch.bfloat16
-            
         model = AutoModelForSequenceClassification.from_pretrained(
             MODEL_NAME, torch_dtype=dtype
         ).to(device).eval()
     return tokenizer, model
 
-THRESHOLD = 0.20
+# Global threshold for the "Verdict" label
+THRESHOLD = 0.60 
 
 # -----------------------------
 # PROTECT STRUCTURE
@@ -82,10 +78,11 @@ def analyze(text):
     blocks = split_preserving_structure(text)
     pure_sents_indices = [i for i, b in enumerate(blocks) if b.strip() and not b.startswith("\n")]
     pure_sents = [blocks[i] for i in pure_sents_indices]
-        
+    
     if not pure_sents:
         return "—", "—", "<em>No sentences detected.</em>", None
 
+    # Context window analysis
     windows = []
     for i in range(len(pure_sents)):
         start = max(0, i - 1)
@@ -101,11 +98,11 @@ def analyze(text):
     weighted_avg = sum(p * l for p, l in zip(probs, lengths)) / total_words if total_words > 0 else 0
 
     # -----------------------------
-    # HTML RECONSTRUCTION
+    # HTML RECONSTRUCTION (Revised Thresholds)
     # -----------------------------
     highlighted_html = "<div style='font-family: sans-serif; line-height: 1.8;'>"
     prob_map = {idx: probs[i] for i, idx in enumerate(pure_sents_indices)}
-
+    
     for i, block in enumerate(blocks):
         if block.startswith("\n") or block.isspace():
             highlighted_html += block.replace("\n", "<br>")
@@ -113,9 +110,14 @@ def analyze(text):
         
         if i in prob_map:
             score = prob_map[i]
-            if score < 0.35: color, bg = "#11823b", "rgba(17, 130, 59, 0.15)"
-            elif score < 0.70: color, bg = "#b8860b", "rgba(184, 134, 11, 0.15)"
-            else: color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)"
+            
+            # Revised coloring logic: < 60% is Human (Green)
+            if score < 0.60: 
+                color, bg = "#11823b", "rgba(17, 130, 59, 0.15)" # Green
+            elif score < 0.80: 
+                color, bg = "#b8860b", "rgba(184, 134, 11, 0.15)" # Amber
+            else: 
+                color, bg = "#b80d0d", "rgba(184, 13, 13, 0.15)" # Red
             
             highlighted_html += (
                 f"<span style='background:{bg}; padding:2px 4px; border-radius:4px; border-bottom: 2px solid {color};' "
@@ -124,10 +126,9 @@ def analyze(text):
             )
         else:
             highlighted_html += block
-
     highlighted_html += "</div>"
-    
-    # Updated Label Logic per your request
+
+    # Final Verdict logic based on the 60% rule
     label = "AI Content Detected" if weighted_avg >= THRESHOLD else "0 or * AI Content Detected"
     
     df = pd.DataFrame({"Sentence": pure_sents, "AI Confidence": [f"{p:.1%}" for p in probs]})
@@ -138,8 +139,8 @@ def analyze(text):
 # -----------------------------
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
     gr.Markdown("## 🕵️ AI Detector Pro")
-    gr.Markdown("Sentence-level analysis with weighted context windows.")
-        
+    gr.Markdown("Sentence-level analysis with weighted context windows. **Threshold: < 60% = Human.**")
+    
     with gr.Row():
         with gr.Column(scale=3):
             text_input = gr.Textbox(label="Paste Text", lines=12)
@@ -147,7 +148,7 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         with gr.Column(scale=1):
             verdict_out = gr.Label(label="Verdict")
             score_out = gr.Label(label="Weighted AI Score")
-            
+    
     with gr.Tabs():
         with gr.TabItem("Visual Heatmap"):
             html_out = gr.HTML()