Yongxin-Guo/trace-uni · Fine tuning trace-uni

I am looking to fine-tune trace-uni on a custom dataset, I want to know the details of sft-youcook2.sh or HFtrainer, which is the better option to finetune trace-uni? I read the paper for trace-uni in which there are details of annotations of dense video captioning task as shown in figure 7 of the TRACE paper.

here token is added 14 times, what is the reason for that? Are they the frames representing the event? What is the video sampling rate?
If I want to fine-tune and use your sft-youcook2.sh , what should be the input format of the data and where should the model checkpoints be located?