Update app.py

app.py

@@ -10,7 +10,7 @@ def predict(user_input):
         result = client.predict(
             user_input, 
             system_prompt,
-            808,    # Max new tokens
+            660,    # Max new tokens
             0.1,    # Temperature
             0.9,    # Top-p (nucleus sampling)
             1.9,    # Repetition penalty