Update app.py

app.py

@@ -163,11 +163,10 @@ def chat_with_codesymb(message: str, history: list) -> str:
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     model.to(device)
 
-    # 🔥 We add a STRICT DIRECTIVE to force it to behave like a chatbot for small talk!
     system_prompt = (
-        "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."
+        "You are CodeSymb, a highly intelligent and empathetic AI coding partner built by "
+        "Developer Dev Malkar and Founder Deepakkumar Malkar at NiftyOracle in India. "
+        "You provide direct, accurate, and clean code. You never repeat yourself. You answer concisely."
     )
     
     prompt = f"<|system|>\n{system_prompt}\n<|user|>\n{message}\n<|assistant|>\n"
@@ -179,6 +178,7 @@ 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
@@ -188,14 +188,19 @@ 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()
 
-            # Clean, stable sampling (No experimental math)
-            logits = logits / 0.6  # Temperature
+            # 1. THE ABSOLUTE REPETITION PENALTY
+            # Mathematically forces the model to pick NEW words instead of repeating
+            for gid in set(generated[-50:]):
+                logits[gid] -= 2.0
 
-            # Top-K
+            # 2. Temperature (Controls creativity)
+            logits = logits / 0.6  
+
+            # 3. Top-K Filtering (Cuts off garbage tokens)
             top_k_values, _ = torch.topk(logits, 40)
             logits[logits < top_k_values[-1]] = -float('inf')
 
-            # Top-P
+            # 4. Top-P Filtering
             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
@@ -208,16 +213,24 @@ def chat_with_codesymb(message: str, history: list) -> str:
             probs = F.softmax(logits_filtered, dim=-1)
             
             next_t = torch.multinomial(probs, 1).item()
+
+            # 5. ANTI-STUTTER MECHANISM
+            # If the model tries to print the exact same token 3 times in a row, kill the loop.
+            if len(generated) >= 3 and generated[-1] == next_t and generated[-2] == next_t:
+                break
+
             generated.append(next_t)
             inp = torch.cat([inp, torch.tensor([[next_t]], device=device)], dim=1)
 
-            # Strict Stop Condition
+            # 6. EXACT STOP SEQUENCES
             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()
 
-    # Clean up output
+    # Clean up any tags that slipped through
     for tag in ["<|user|>", "<|system|>", "<|assistant|>", "<|endoftext|>"]:
         response = response.split(tag)[0]