Update app.py

app.py

@@ -163,11 +163,11 @@ def chat_with_codesymb(message: str, history: list) -> str:
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     model.to(device)
 
-    # I slightly shortened the prompt to make it heavily focused on answering directly.
+    # 🔥 We add a STRICT DIRECTIVE to force it to behave like a chatbot for small talk!
     system_prompt = (
-        "You are CodeSymb, a deeply empathetic and highly intelligent AI coding partner. "
-        "You were built by Developer Dev Malkar and Founder Deepakkumar Malkar at NiftyOracle in India. "
-        "Always answer the user's questions directly, politely, and accurately."
+        "You are CodeSymb, an intelligent and empathetic AI coding partner built by Developer Dev Malkar and Founder Deepakkumar Malkar at NiftyOracle in India. "
+        "DIRECTIVE: If the user just says 'Hi', greets you, or asks about your creators, you MUST answer in 1 or 2 normal sentences WITHOUT writing any code. "
+        "ONLY write code if the user explicitly asks for code."
     )
     
     prompt = f"<|system|>\n{system_prompt}\n<|user|>\n{message}\n<|assistant|>\n"
@@ -179,7 +179,6 @@ def chat_with_codesymb(message: str, history: list) -> str:
     inp = torch.tensor([prompt_ids], dtype=torch.long).to(device)
 
     END_SEQ  = sp.encode("<|endoftext|>")
-    USER_SEQ = sp.encode("<|user|>")
     
     generated = []
     MAX_NEW = 512
@@ -189,24 +188,14 @@ def chat_with_codesymb(message: str, history: list) -> str:
             with torch.autocast("cuda" if device.type == "cuda" else "cpu", dtype=torch.bfloat16):
                 logits = model(inp)[0, -1, :].float()
 
-            # 1. Temperature (Increased to 0.7 to stop repetitive stuttering)
-            logits = logits / 0.7
+            # Clean, stable sampling (No experimental math)
+            logits = logits / 0.6  # Temperature
 
-            # 2. Repetition Penalty (Applied to ALL generated tokens so it never repeats sentences)
-            penalty = 1.3
-            for tok_id in set(generated):
-                if logits[tok_id] > 0:
-                    logits[tok_id] /= penalty
-                else:
-                    logits[tok_id] *= penalty
+            # Top-K
+            top_k_values, _ = torch.topk(logits, 40)
+            logits[logits < top_k_values[-1]] = -float('inf')
 
-            # 3. Top-K Filtering (Cuts off hallucinated garbage tokens)
-            top_k = 40
-            top_k_values, _ = torch.topk(logits, top_k)
-            min_top_k = top_k_values[-1]
-            logits[logits < min_top_k] = -float('inf')
-
-            # 4. Top-P Filtering
+            # Top-P
             sorted_logits, sorted_idx = torch.sort(logits, descending=True)
             cum_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
             remove = cum_probs > 0.90
@@ -222,15 +211,13 @@ def chat_with_codesymb(message: str, history: list) -> str:
             generated.append(next_t)
             inp = torch.cat([inp, torch.tensor([[next_t]], device=device)], dim=1)
 
-            # Check if it generated the stop sequence
+            # Strict Stop Condition
             if len(generated) >= len(END_SEQ) and generated[-len(END_SEQ):] == END_SEQ:
                 break
-            if len(generated) >= len(USER_SEQ) and generated[-len(USER_SEQ):] == USER_SEQ:
-                break
 
     response = sp.decode(generated).strip()
 
-    # Strip tags to leave clean text
+    # Clean up output
     for tag in ["<|user|>", "<|system|>", "<|assistant|>", "<|endoftext|>"]:
         response = response.split(tag)[0]