Why do I get very low initial bbox values when using the COCO dataset to fine-tune SAM3？

Why do I get very low initial bbox values when using the COCO dataset to fine-tune SAM3？
The pre-trained SAM3 should have relatively high detection capability on the COCO dataset, shouldn't it?

No file, test, command, configuration, or reproducible example is provided; the issue only includes an image and a question about COCO fine-tuning. Start by collecting the fine-tuning setup and comparing the initial bounding-box outputs with the COCO annotations and pretrained SAM3 behavior. Done means identifying and documenting the cause of the low values.

Written by the indexing model from the issue text.