RT Detr in ultralytics accuracy not reproducible

[Mythili-kannan]

I trained a custom dataset of RT-DETR architecture using ultralytics repo, its works good and the detection results are reliable,

when i tried to train the same dataset, and convert the txt to json and train here, i am unable to get same reliable results here, its having so much false predictions and confidence score are poor

can you please help me to solve this? Am I missing something here