astronolan/hsc-anomaly-expert

HSC Anomaly Expert

This dataset contains 1,000 galaxy images independently scored by 16 expert astronomers. The ten fixed few-shot examples form the train split; the remaining 990 images form the test split used for evaluation, containing 45 expert-consensus anomalies and 945 other images.

The images come from public data release 2 of the Subaru Hyper Suprime-Cam (HSC) survey; this is the HSC in the dataset name. The source images, Zooniverse subject metadata, volunteer measurements, and mantha_ml_anomaly_score come from Mantha et al. (2024), Through the citizen scientists' eyes: Insights into using citizen science with machine learning for effective identification of unknown-unknowns in big data, Citizen Science: Theory and Practice, 9(1), Article 40, 1–15. Please cite Mantha et al. when using this dataset.

expert_consensus_anomaly is true when at least 40% of the 16 experts selected an image. The accompanying count, total, and fraction columns are aggregate survey measurements; individual expert identities and votes are not included. volunteer_selected_fraction is the source table's percent_selected divided by 100, and mantha_ml_anomaly_score is its anomaly_score.

Images were downloaded from the Zooniverse URLs recorded by Mantha et al. and stored byte-for-byte as their original PNGs, without resizing or re-encoding. Metadata were joined by the numeric image filename, and rows retain the filename-sorted order used to construct the evaluation set.