Lower for testing

app.py

@@ -33,7 +33,7 @@ def apply_tiny_chat_template(messages):
             prompt += f"<|response|> {content}\n"
     return prompt
 
-def chat_fxn_caller(message, history, system_prompt=system_prompt_default, temperature=0.6, top_p=0.95, max_tokens=256):
+def chat_fxn_caller(message, history, system_prompt=system_prompt_default, temperature=0.6, top_p=0.95, max_tokens=128):
     messages = []
     if system_prompt.strip():
         messages.append({"role": "system", "content": system_prompt})
@@ -50,7 +50,7 @@ def chat_fxn_caller(message, history, system_prompt=system_prompt_default, tempe
         prompt,
         do_sample=True,
         max_new_tokens=max_tokens,
-        top_k=50,
+        top_k=30,
         temperature=temperature,
         num_return_sequences=1,
         repetition_penalty=1.1,