DETRv2 training notebook produces bad model

[matheger]

Trying to run through https://github.com/qubvel/transformers-notebooks/blob/main/notebooks/RT_DETR_v2_finetune_on_a_custom_dataset.ipynb at commit 89a48c3

The trained model performs much worse than the output cells suggest. Best detection score in the last inference step is ~0.2 or less.

I had to change a few things to make this notebook work (these shouldn't really affect the model):

1. Drop three examples with bad bounding boxes from dataset["train"] (indices 65, 383, 476)
2. Fix the category issue from here
3. Set dataloader_num_workers=0, otherwise they keep crashing (multiprocessing issues?)
4. Save to/load from disk instead of pushing to hub

My environment:

python 3.11
transformers 4.56.2
torch 2.8.0