Confusion regarding rotated annotations

[adityasihag1804]

I'm trying to do rotated detections on text pages. It is a custom dataset. I'm confused regarding the bbox values and segmentation values. here's a sample image from the training dataset, with a few axis-aligned (Blue), also the minimum spanning rectangle, and rotated boxes (green).

So for the field bbox , x y w h , I have tried passing the values for axis-aligned and rotated both.
And for segmentation , I'm always passing the 4 coordinates of the green box.
But the model seems to not be learning anything. The loss seems to be decreasing, but mAP remains 0 the entire time. Please let me know if I'm doing something wrong here.