Question about the “support_docs” field in data processing for step reward

[ZijunSong]

Hi, thank you for sharing such great work! I’m very interested in applying your distributed reward mechanism in my research.

While going through the data processing code, I noticed that only the data_process/musi_search dataset includes the "support_docs" field (from example['sub_support_docs']), while the other two datasets seem to have this field empty.

May I ask if this is because the current implementation only supports that dataset? And if I’d like to use your method on other datasets, should I manually provide the ground-truth reference documents for "support_docs"?

[zxh20001117]

Currently, our training set only includes the MusiQue dataset and its supporting documentation (golden evidence). If you need to train using other datasets or custom datasets, you will need to prepare your own support documents according to the corresponding format.