Directly finetuning without Mrefm on the single dataset finetuning setting

[ShuimuZeng]

I tried to replicate the results of table 15 in the oneref paper. However, I cannot achieve the same results in the the table for the baseline on single dataset finetuing setting. I used the script from the train_rec_single_dataset_finetuning_base.sh and directly finetunes from the BEiT-3 parameters for the single dataset finetuing setting. Without any augmentation for the REC forward logic, my result is higher that the baseline mentioned in table 15. I also noticed that the training details showed in the paper is not the same as in the shell scripts. I wonder if the authors could clarify:

1. What training protocol was used for the baseline?
2. Has the training
3. How to reproduce the 85.23% 88.13% 83.82% result for the single dataset finetuing setting of the baseline in table 15?