Derived Face\u2013Hand Corpus from HaGRID

This dataset is a derived version of the HaGRID corpus, curated to enable unified gender prediction from either hand or face images. The original HaGRID dataset contains over 550,000 gesture images; we retain only palm-visible gestures (palm, stop, stop_inverted) to ensure unobstructed hand regions and frequent facial visibility. Each image is automatically cropped into aligned hand and face regions, followed by cleaning, alignment, and construction of subject-exclusive train\/validation\/test splits. The resulting corpus comprises 15,064 paired hand\u2013face images. Labels correspond to apparent gender, obtained through a semi-supervised pipeline that combines original metadata, FairFace-based predictions, and manual review. Demographic metadata (age, apparent race) are included strictly for fairness evaluation and not as model inputs. The dataset is released together with a snapshot of the original HaGRID and its license, supporting reproducible research in cross-modality biometrics, gender prediction, and fairness analysis.