Rag Prompt

#67
by devve1 - opened

Hello, i was trying the model on my rag recently. I juste rewrote by saying "write no_answer" the very classical prompt on langchain hub :
"""You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you don't know the answer, just write 'no_answer'. Use three sentences maximum and keep the answer concise.

Question: {question}

Context: {context}

Answer:"""

and it always reply with 'no_answer'

I took the very same example you have put on your model card with vLLM, to reproduce that.
Do you have any idea why its replying everytime "no_answer" ?

N.B : i was trying llama 3 8b on the same prompt back then with the same setup ( and same context lenght of 8192 ) , and it was working fine

EDIT : your prompt in your needle of the haystack test is the only one that work : {context} {query} Don't give information outside the document or repeat your findings. Keep your response short and direct.

but if i try to add to this prompt to tell the ai to not answer when it doesn't know, it always says it doesn't know

This comment has been hidden

Sign up or log in to comment