Fine-tunning model on instance segmentation tasks without prompts

[JIBSN]

Hi,
I want to fine tune the seem_focall_v1 model on a custom coco style dataset for instance segmentation task with only a few classes. I currently have no need to do interactive segmentation and just want to do instance segment. I have prepared a dataset and registed it, but I can't train it. I found there are some errors when calculate loss. Could you please give me some tips on train the model without prompts?
Thanks