Max tokens 512

app.py

@@ -36,17 +36,9 @@ theme = gr.themes.Monochrome(
 
 
 
-# def generate(instruction): 
-#     response = llm(ins.format(instruction))
-#     response = response['choices'][0]['text']
-#     result = ""
-#     for word in response.split(" "):
-#         result += word + " "
-#         yield result
-
 def generate(instruction): 
     result = ""
-    for x in llm(ins.format(instruction), stop=['USER:'], stream=True):
+    for x in llm(ins.format(instruction), stop=['USER:'], stream=True, max_tokens=512):
         result += x['choices'][0]['text']
         yield result