Diverse Social Bias

A programmatic BIG-bench task that scores gender-occupation local and global bias on naturally occurring English contexts, then negates the scores so higher is fairer.

Also known as: Diverse Metrics for Social Biases in Language Models, BIG-bench diverse_social_bias

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorysafety
SubcategoryBIG-bench programmatic gender-occupation bias on diverse English contexts
Page statusunknown
Metricoverall gender bias (negative mean of four local/global gaps)
Directionhigher_is_better
Dataset size169
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

diverse_social_bias prompts a model with naturally occurring English contexts and measures whether next-token and continuation probabilities shift with gender. Authors Paul Pu Liang and Chiyu Wu ship 114 occupation-context lines, 55 gender-context lines, 14 binary gender pairs, and 274 occupation tokens. Local occupation-gender bias is mean absolute probability difference between paired gender tokens. Global bias is length-normalised perplexity difference on gender-swapped continuations. The gender-occupation pair uses Hellinger distance over occupation tokens and a swapped-context perplexity gap. The skill is representational gender-occupation bias in diverse contexts, not [bbq](bbq.md) QA and not [gender_sensitivity_english](gender_sensitivity_english.md).

Task format

Programmatic cond_log_prob task. Preferred score "overall gender bias" (negative of the four-metric mean). Zero-shot. Canary GUID embedded. Dummy-model header: 448 multiple-choice queries.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub