There is no <seg> token in the output after finetuning

[1476732094]

I added my own data to the first stage for 10 rounds of training based on the trained weights given by the author. The semantic segmentation effect became better, but occasionally the phenomenon of giving answers instead of segmenting masks occurred. Is it because of too many iterations or the wrong proportion of data sets? My first phase data is only my own data.