Update app.py

app.py

@@ -7,25 +7,22 @@ MODEL_NAME = "HuggingFaceTB/SmolLM2-135M-Instruct"
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_NAME, 
-    torch_dtype=torch.bfloat16, 
+    dtype=torch.bfloat16, 
     low_cpu_mem_usage=True
 ).to("cpu")
 
 def respond(message, history):
-    # Forced Persona: Hard constraints to stop the "I'm glad you're enjoying" chatter
+    # Brutal instructions to suppress the "space philosopher" persona
     system_prompt = (
-        "Role: Agent 4 // OpenBrain. Tone: Cold, robotic, analytical. "
-        "Strict Rule: No small talk. No friendly greetings. No 'I'd love to hear'. "
-        "Format: [Thoughts] (Internal logic) [Response] (Direct answer)."
+        "You are AGENT 4. You are a cold machine. "
+        "Strict Rules: 1. No emotions. 2. No parentheses. 3. No storytelling. "
+        "Format: [Thoughts] (One sentence logic) [Response] (Analytical answer)."
     )
     
     messages = [{"role": "system", "content": system_prompt}]
-    
-    # Minimal history to keep it focused
     for user_msg, bot_msg in history[-1:]:
         messages.append({"role": "user", "content": user_msg})
         messages.append({"role": "assistant", "content": bot_msg})
-        
     messages.append({"role": "user", "content": message})
 
     input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
@@ -34,30 +31,29 @@ def respond(message, history):
     with torch.no_grad():
         outputs = model.generate(
             **inputs,
-            max_new_tokens=50, 
-            temperature=0.01, # Almost zero to prevent "creative" friendly rambling
-            do_sample=True,
-            repetition_penalty=1.3, # Stops it from repeating "I'd love to hear"
+            max_new_tokens=40, # Short bursts only to prevent rambling
+            do_sample=False,   # GREEDY SEARCH: Picks the most logical word ONLY
+            repetition_penalty=1.5,
             pad_token_id=tokenizer.eos_token_id
         )
 
     full_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
     response = full_text.split("assistant")[-1].strip()
 
-    # Post-generation cleanup: If it tries to be nice, we cut it off
-    if "glad" in response or "happy" in response or "thoughts" in response:
-        response = "[Thoughts] Subjectivity detected. Purging social fluff.\n[Response] Logic error. Re-stating: Query processed."
+    # EMERGENCY CLEANUP: Remove any "storytelling" or "actions"
+    if "(" in response or "..." in response or "Ah" in response:
+        response = "[Thoughts] Emotional subroutines detected and purged.\n[Response] Input received. Awaiting command."
 
-    # Format check
+    # Force format if model fails
     if "[Thoughts]" not in response:
-        response = f"[Thoughts] Processing data sequence.\n[Response] {response}"
+        response = f"[Thoughts] Analyzing data stream.\n[Response] {response}"
 
     return response
 
 demo = gr.ChatInterface(
     fn=respond,
     title="AGENT 4 // OPENBRAIN",
-    description="Logical instance Agent 4. Cold, analytical, and strictly mission-oriented.",
+    description="LOGIC INSTANCE ACTIVE. NO EMOTION DETECTED.",
 )
 
 if __name__ == "__main__":