A question about the use of "qa_id" of each dataset item about mac HOT 2 CLOSED

Hi, thank you for your interest in our paper!

This is a typical way to define QA loss in the literature, and it is unrelated to online adaptation. Specifically, we use teacher forcing, which is commonly used for autoregressive modeling (using the ground truth previous tokens to predict the next token). In the QA setup, it is using the previous answer token (and the question tokens) to predict the next answer token.

Also, note that this is used for the training stage, hence, there is no leakage in the online adaptation stage (we do not use the answer when adapting and generating the answer).

• https://github.com/jihoontack/MAC/blob/main/evals/qa_utils.py#L91
• This part generates the answer where we do not use the answer.

We are still refactoring and polishing the code, so if you have any bugs or issues, please let us know!

Thank you again for your interest.

Best,
Jihoon

from mac.

Thanks for your kind help. I have another question: what is the "lift_ratio" in the file config?

from mac.