mlnomad/imagenet-1k-224-less

ImageNet-1k-224 LESS

LESS (Lightweight Embedding-driven Subset Selection) coreset of mlnomad/imagenet-1k-224.

How it was built

1. CLIP (google/siglip-base-patch16-224) embeddings were generated for every image in the train split.
2. For each of the 1,000 ImageNet classes, k-means clustering (k=10) was applied to the class embeddings.
3. The image closest to each centroid was selected, yielding at most k images per class — the most geometrically representative samples.

Columns

Column	Type	Description
image	PIL Image	224×224 RGB image
label	int (0–999)	ImageNet class index
original_index	int	Row index in the source dataset for cross-referencing

Stats

• Source: mlnomad/imagenet-1k-224 (train split)
• k (images per class): 10
• Total images: 10000
• CLIP model: google/siglip-base-patch16-224

Cross-reference with the source dataset via the original_index column.