quinnlue/

imagenet100-lt-16k

A self-contained, deterministic long-tail subset of clane9/imagenet-100, built as a fixed constant for fair comparison across experiments.

Split Rows Notes

train 16,000 exponential long tail, imbalance factor 100

validation 2,000 20 per class, from the official val split

test 2,000 20 per class, disjoint from validation

All images are RGB, resized (shorter side -> 256) and center-cropped to 256x256, stored as JPEG q95.… See the full description on the dataset page: https://huggingface.co/datasets/quinnlue/imagenet100-lt-16k.