optimized hyperparams: cosine LR, copy_paste, augmentation tuning, 100 epochs

fine-tune from best.pt, 75 epochs, drop drone images

set epochs to 200 for full convergence run

bump to 150 epochs and add test split for final evaluation

reduce epochs to 40 for small dataset

increase batch to 128 and imgsz to 1280 for full VRAM utilization

model update

initial commit