Update app.py

app.py

@@ -26,8 +26,8 @@ def smart_response(message, history):
 
     if is_opinion:
         print(f"🧠 OPINION MODE: {message}")
-        # DUAL FRAME PROMPT (Fixed for stability)
-        # We ensure it stops generating after the response
+        # DUAL FRAME PROMPT
+        # We use the standard instruction format the model learned during fine-tuning.
         prompt = f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.
 
 ### Instruction:
@@ -35,20 +35,25 @@ def smart_response(message, history):
 
 ### Response:
 """
-        # Added double newline to stop early if it tries to start a new header
-        stop_tokens = ["</s>", "###", "\n###"]
         
     else:
         print(f"ℹ️ CHAT MODE: {message}")
-        # STANDARD CHAT PROMPT (Fixed)
-        # 1. Added '<s>' (Beginning of String) - CRITICAL for Mistral
-        # 2. Removed "Question:" label (This prevents the 'Quiz Mode' loop)
-        # 3.Kept the structure simple so it knows it is a chat
-        prompt = f"<s>[INST] {message} [/INST]"
-        
-        stop_tokens = ["</s>", "[INST]"]
+        # FACTUAL PROMPT
+        # CRITICAL FIX: We use the SAME format (Alpaca), but we add a specific
+        # instruction telling the model to be "concise" and "factual".
+        prompt = f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+You are a helpful assistant. Answer the following question concisely and directly: {message}
+
+### Response:
+"""
 
     # --- GENERATION ---
+    # We add "###" to the stop tokens. This cuts off the model immediately 
+    # if it tries to start generating "Solution 1" or a new "### Instruction".
+    stop_tokens = ["</s>", "###", "Solution"]
+
     output = llm(
         prompt,
         max_tokens=512,