Captioned bucketed export of timm/imagenet-1k-wds.
Images are grouped into precomputed aspect-ratio buckets and stored in uncompressed tar shards for efficient loader-time batching. The export uses a 512 -> 384 -> 256 base-resolution waterfall: each image is assigned to the highest bucket family it fits without upsampling.
This export uses image_mode=passthrough_jpeg, so the stored .jpg bytes are taken directly from the source dataset. The exporter… See the full description on the dataset page:
https://huggingface.co/datasets/data-archetype/imagenet_1k_bucketed_captioned.
