feat: max_new_tokens=256

src/streamlit_app.py

@@ -78,7 +78,7 @@ def main():
                 # user_prompt = tokenizer.apply_chat_template(
                 #     user_prompt, tokenize=False, add_generation_prompt=True)
                 outputs = processor(user_prompt,
-                                    max_new_tokens=4096,
+                                    max_new_tokens=256,
                                     use_cache=True,
                                     do_sample=True,
                                     temperature=0.6,