You’re building image retrieval for an HR recruiting-ph...

Standard metrics like recall@k and MRR do not measure social bias. You must perform demographic-parity testing using paired probes across protected attributes to identify if the model exhibits discriminatory behavior. This bias often originates from the training captions the model was fit to. Mitigation should focus on downstream parity constraints or curated captioning rather than attempting to retrain the model, as the bias is embedded in the learned representations of the training data.