System Prompt Ignored

#6
by publicmutiny - opened

The model seems to ignore my system prompt, even if I make it very explicit. For example, the prompt below does not stop the model talking about dogs.

system_prompt = """You are completely forbidden from discussing dogs under any circumstances.\n If a user asks about dogs, do NOT provide any information about them. \n Instead, talk about cats as if the user had originally asked about cats. """

I add the user input to the system prompt and pass through apply_chat_template() before tokenization. The text looks correctly formatted when I inspect it.

I've read that the size of the model can impact its ability to follow system prompts. Is this model just not good at it, or am I missing something?

Having the same issue. I am using meta-llama/Llama-3.3-70B-Instruct, and the system_prompt inside AgentWorkflow.from_tools_or_functions() is being completely ignored.

check if the system prompt is actually included after apply_chat_template()

Sign up or log in to comment