Explore high-quality datasets for your AI and machine learning projects.
CrowS‑Pairs is a challenging dataset for evaluating social bias in masked language models. It contains 1,508 test samples, each comprising two sentences—one more biased and one less biased. The dataset covers various bias types such as race, gender, religion, etc. It was constructed from the fictional portions of ROCStories and MNLI and annotated via crowdsourcing.
