PaliGemma Debiasing Methods Comparison

Compare all debiasing methods side by side. Lambda controls Logit Adjustment strength (default: 5.0). VDD Lambda controls Visual Debias Decoding strength (default: 1.0). Threshold controls Logit Adjustment activation. Choose 'Alignment' (s_scale = classifier output) or 'Neutralization' (s_scale = 0) mode for Logit Adjustment.

0 100
0 1
s_scale Mode

Alignment: s_scale = classifier output (gender prediction). Neutralization: s_scale = 0.

0 10