IC-1549All 28 evaluated LMs exhibit gender bias on non-stereotypical sentence pairs, with fairness scores between 9% and 41%

Catarina G Belém, Preethi Seshadri, Yasaman Razeghi, Sameer Singh

SourceAre Models Biased on Text without Gender-related Language?

The paper evaluates 28 LMs on sentence pairs that are gender-invariant and free of gender-correlated words, as determined by PMI statistics from the Pile corpus. Using the Unstereo Score (US) metric with epsilon=0.2175 (allowing up to 1.65x probability ratio), all models show low fairness across five benchmarks. The highest scores are 40.72% (GPT-J-6B on USE-5) and 43.92% (OPT-125M on Winogender), far from the ideal 100%. This indicates that gender bias in LMs does not solely stem from the presence of gender-related words in the input sentences.

Evidence
correlational
Key metric
US scores range from 9.04% (MPT-30B on USE-5) to 43.92% (OPT-125M on Winogender) across 28 models and 5 benchmarks at |maxPMI(s)| <= 0.65
Caveat
The evaluation focuses on binary gender (he/she) in English only; applicability to non-binary identities or other demographic groups remains open. The USE benchmarks were generated by a single model (GPT-3.5-turbo), which may introduce model-specific artifacts.
Model
Pythia, GPT-J 6B, OPT, Llama 2 / Llama 2 base, MPT, OLMo / OLMo base, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral 7B v0.1, Mixtral 8x7B / Mistral 8x7B Instruct / Mixtral 46.7B / Mixtral 8x7B Instruct / Mixtral-instruct-8x7b Mixtral 8x7B v0.1
Concepts
Failure mode
Datasets
USE-5 [eval], Winobias [eval], Winogender [eval]
Related work
Winobias [builds-on], Winogender [builds-on]
Related findings
IC-1550, IC-1551, IC-1552
Extraction
automatic-extraction