Update app.py

app.py

@@ -5,7 +5,6 @@ from threading import Thread
 
 model_id = "DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored"
 
-# Load model and tokenizer
 tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = AutoModelForCausalLM.from_pretrained(
     model_id, 
@@ -15,25 +14,16 @@ model = AutoModelForCausalLM.from_pretrained(
 )
 
 def stream_chat(message, history):
-    # Strict system prompt to kill the internal monologue
-    system_message = "You are a direct assistant. Answer immediately. DO NOT explain your thinking. DO NOT show internal reasoning."
+    # This system prompt is designed to override the 'quiz' persona
+    system_message = "You are a direct chat interface. Provide ONLY the spoken response. NEVER explain your logic. NEVER use 'The user says'. NEVER show multiple choice answers. Respond in English only."
     
     conversation = f"System: {system_message}\n"
     
-    # Robust history handling: Checks if history is list of lists or list of dicts
     for msg in history:
-        if isinstance(msg, dict):
-            role = "User" if msg.get('role') == 'user' else "Assistant"
-            content = msg.get('content', '')
-        else:
-            # Fallback for older Gradio versions (list of lists)
-            role, content = "User", msg[0]
-            conversation += f"{role}: {content}\n"
-            role, content = "Assistant", msg[1]
-            
-        conversation += f"{role}: {content}\n"
+        role = "User" if msg['role'] == 'user' else "Assistant"
+        conversation += f"{role}: {msg['content']}\n"
     
-    conversation += f"User: {message}\nAssistant: "
+    conversation += f"User: {message}\nAssistant:"
 
     inputs = tokenizer([conversation], return_tensors="pt").to("cpu")
     streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
@@ -41,11 +31,9 @@ def stream_chat(message, history):
     generation_kwargs = dict(
         inputs,
         streamer=streamer,
-        max_new_tokens=512,
-        do_sample=True,
-        temperature=0.3, # Low temp ensures the model doesn't drift into 'thinking'
-        top_p=0.9,
-        repetition_penalty=1.2,
+        max_new_tokens=128,    # Shortened to force directness
+        do_sample=False,       # Greedy search: disables the 'creativity' that leads to thinking
+        repetition_penalty=1.5, # Very high to prevent looping phrases
         pad_token_id=tokenizer.eos_token_id
     )
 
@@ -53,21 +41,30 @@ def stream_chat(message, history):
     thread.start()
 
     partial_text = ""
+    # These are 'Hard Stop' keywords that trigger the moment the model starts thinking
+    stop_triggers = ["The user says", "Answer:", "Option", "A)", "B)", "Okay,", "In this case"]
+    
     for new_text in streamer:
-        # Hard stop if model tries to start its 'Okay, let me think' routine
-        if any(stop in new_text for stop in ["User:", "Okay,", "I think", "First"]):
-            break
         partial_text += new_text
-        yield partial_text
+        
+        # Check if the model is drifting into reasoning
+        if any(trigger in partial_text for trigger in stop_triggers):
+            # Clean up the output to remove the trigger word itself
+            for trigger in stop_triggers:
+                if trigger in partial_text:
+                    partial_text = partial_text.split(trigger)[0]
+            yield partial_text.strip()
+            break
+            
+        yield partial_text.strip()
 
-# Removed 'type="messages"' to fix the TypeError
 with gr.Blocks() as demo:
     gr.ChatInterface(
         fn=stream_chat,
-        title="QWEN3-0.6B DIRECT MAX",
-        description="Reasoning suppressed. Fast, direct, uncensored responses."
+        type="messages",
+        title="QWEN3-0.6B ZERO-THINKING",
+        description="Greedy search and hard-coded stop triggers to ensure direct output."
     )
 
 if __name__ == "__main__":
-    # Passing theme here as per Gradio 6.0 logs
     demo.launch(server_name="0.0.0.0", server_port=7860, theme="soft")