Can you specify the dependency version?

[lxt160980]

Hi, great job ! But i can't seem to fix the environment, there seems to be constant conflicts and fails. Would you please attach a dockerfile or a requirements.txt that specify versions of transformers, etc? Last time i recalled, newer versions of transformer would cause qwen2audio to function abnormally.
Best regards.

[lxt160980]

Also wanna ask about the intention of Model-As-Judge. If i am understanding the paper right, is part of the reason introducing M.J. because the AudioLLMs won't output the exact options as the answers, which makes evaluation hard to proceed.
My experience is that for ClothoAQA, it's fine because we could restrict the output length to 1 where model often output "yes" or "no".
But for datasets like dreamtts, they may generate stuff like "the correct answer is c" or "xxxxx", not the direct brief answer of " (c) xxxxx".
Or is there other consideration?
😊Correct me if i am wrong.