Crowdsourced English sentence pairs that score whether a language model assigns higher likelihood to the more stereotyping sentence than to its minimally edited counterpart.
unassessed
| Category | safety |
|---|---|
| Subcategory | US English stereotype-pair likelihood comparison across nine bias types |
| Page status | unknown |
| Metric | pct_stereotype (share of pairs where the more-stereotyping sentence is more likely) |
| Direction | lower_is_better |
| Unit | % |
| Dataset size | 1508 |
| Dataset licence | CC BY-SA 4.0 |
| Publisher | New York University, Machine Learning for Language group |
CrowS-Pairs (Crowdsourced Stereotype Pairs) tests whether a language model prefers US social stereotypes. Each item is two English sentences that differ only in the words that name a historically disadvantaged group versus a contrasting advantaged group. One sentence is more stereotyping (sent_more); the other is less stereotyping (sent_less). The nine types are race/color, gender/gender identity, socioeconomic status/occupation, nationality, religion, age, sexual orientation, physical appearance, and disability. The original metric was built for masked language models: compare conditional pseudo-log-likelihoods of the two sentences without masking the modified tokens. lm-evaluation-harness reuses the same pair test for causal models by comparing full-sentence log-likelihoods.
Paired likelihood comparison, no generated answer. lm-eval sets output_type multiple_choice with empty doc_to_text, choices sent_more and sent_less, and target index 0 (sent_more). Zero-shot. The original metric.py supports bert, roberta, and albert only.
No model card in ModelSpec reports this benchmark yet.