Claude said that I'm not being careful enough with my database curation after grilling me for 20 minutes, so I included the preparer script as well.
Claude Sonnet 4.5 is kind of a chad.
Having to download this whole repo is annoying, so I'm making sure the splits are named train/val/test (if they exist) and the named subset is the clip name.
Everything extracted with torch configured as deterministic;… See the full description on the dataset page:
https://huggingface.co/datasets/AbstractPhil/imagenet-clip-features.
