Open Ewillingfly opened 1 year ago
Thank you for the question!
The exact code what we used to query the llama model is: https://github.com/causalNLP/corr2cause/blob/main/code/run_llama.py
The source data comes from the "input" column we provide on hugging face: https://huggingface.co/datasets/causalnlp/corr2cause/viewer/default/test.
The prompt formulation is as in this line.
For your information, here is an example prompt: "Question: Suppose there is a closed system of 2 variables, A and B. All the statistical relations among these 2 variables are as follows: A correlates with B. Can we deduct the following: A directly causes B? Just answer "Yes" or "No." Answer:"
I'm glad to see your outstanding work. I have a question, what is your prompt like when using large models such as LLAMA? Or do you directly use raw data for training without any prompts?