09 WAAM-ViD HF Format
videos.zip contains the original MP4 files only. model_checkpoint/ is
excluded. Train/test image frames are converted into the shared SFT schema:
annot.label is the per-frame Width_mm value and annot.annotation is the
COCO melt-pool semantic segmentation, or the string nan when no segmentation
annotation exists for that image.