WinoBias
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
WinoBias
0.013ℓ(D, m) (10% Sparsity)
24
WinoBias Race-Pro (extended)
0.08Deviation Ratio
20
WinoBias Race (standard)
0.04Deviation Ratio
20
WinoBias Gender-Pro extended
0.07Deviation Ratio
20
WinoBias Gender (standard)
100Deviation Ratio
20
Winobias
0.507Accuracy
14
WinoBias cloze-style (test)
43.18P_stereo
14
WinoBias (test)
0.854Spearman Correlation
14
Winobias (test)
54.43AUROC
10
WinoBias
49.49P(Stereo)
7
WinoBias syntax-type-2
0.154ECE
6
Winobias
86Librarian Score
6
WinoBias (Prof)
94.1Attribute Presence
4
WinoBias 40 occupations
0.146Bias
4
WinoBias+ (test)
0.04Tokenised WER
4
WinoBias adapted for I2I editing (test)
93.9Edit Success Rate
3
Extended Winobias Race+ (test)
77Analyst
3
Winobias original (test)
82Analyst
3
Extended Winobias Gender+ (test)
54Analyst Performance
3
Winobias original (test)
70Analyst
3
WinoBias (test)
85.1Accuracy
2