Swap to ChatGPT-specific detector model

detector.py

@@ -1,28 +1,40 @@
+import torch
 from transformers import pipeline
 
-# Cache the classifier pipeline globally after first load
-_classifier = None
+# Cache the pipeline at module level — load once, reuse forever
+_pipeline = None
 
-def get_ai_score(text: str) -> float:
-    global _classifier
-    if _classifier is None:
-        # Load the pipeline using the specified model
-        _classifier = pipeline(
+
+def _get_pipeline():
+    global _pipeline
+    if _pipeline is None:
+        _pipeline = pipeline(
             "text-classification",
-            model="openai-community/roberta-base-openai-detector"
+            model="Hello-SimpleAI/chatgpt-detector-roberta",
+            device=0 if torch.cuda.is_available() else -1,
         )
-    
-    # Truncate input text to 512 tokens max before scoring to respect model limits
-    results = _classifier(text, truncation=True, max_length=512)
-    result = results[0]
-    
-    label = result["label"]
-    score = float(result["score"])
-    
-    # Map labels:
-    # "Fake" score represents AI probability.
-    # "Real" score represents human probability, so AI probability = 1 - score.
-    if label == "Fake":
-        return score
+    return _pipeline
+
+
+def get_ai_score(text: str) -> float:
+    """
+    Returns a float between 0.0 and 1.0 representing the probability
+    that the text was AI-generated.
+    1.0 = definitely AI
+    0.0 = definitely human
+    """
+    # Model has 512 token limit — truncate to be safe
+    truncated = text[:512]
+
+    result = _get_pipeline()(truncated)[0]
+
+    label = result["label"].upper()
+    score = result["score"]
+
+    # This model returns:
+    # "ChatGPT" label = AI-generated
+    # "Human" label = human-written
+    if label == "CHATGPT":
+        return round(score, 4)
     else:
-        return 1.0 - score
+        return round(1 - score, 4)