Three small, clean text-classification datasets that unrelabel uses to show off data-poisoning attacks and the defenses that catch them. Small on purpose: the whole attack, retrain, and measure loop runs in milliseconds on a laptop, no GPU, no waiting around.
These are the clean baselines. The poisoned versions live in the companion poison benchmark.
ecommerce-review-sentiment (1810 train / 452 test) classes: negative (915), positive (895)
sms-spam (4459… See the full description on the dataset page:
https://huggingface.co/datasets/o22y/unrelabel-demos.
