TransferZ: a photometric dataset for machine learning

Overview

TransferZ is a machine learning ready dataset containing Hyper-Suprime Cam (HSC) PDR2 grizy photometry and COSMOS2020 derived photometric redshifts for 116,335 galaxies in the 2 sq. deg. COSMOS field. The dataset is associated with the paper "Improving Generalization and Uncertainty Quantification of Photometric Redshift Models" by Soriano et al. (2025). It is designed for machine learning applications in astrophysics, particularly for redshift estimation. Soriano et al. (2025) used TransferZ complimentary to GalaxiesML (doi) to test methods of generalizing redshift models. We provide the same train test splits used in the paper. In addition, we provide a "conformal" set which was used in the application of conformal prediction. 

Features

Photometry for 116,335 galaxies in five photometric bands (g,r,i,z,y)

Photometric redshifts for each galaxy derived from 35-band photometry (see Weaver+22). Redshifts range from 0.01 to 4

Citations

If you make use of any of these products, please cite this repository and Soriano et al. (2025). In addition, for COSMOS2020 data products, please cite Weaver et al. (2022); for HSC PDR2 data, please cite Aihara et al. (2019); for GalaxiesML data products, please cite Do et al (2024).

References:

Aihara et al. (2019). PASJ, 71, 6. [arXiv:1905.12221][doi]

Weaver et al. (2022).ApJS, 258, 11. [arXiv:2110.13923][doi]

Do et al. (2024). [arXiv:2410.00271]

Soriano et al. (2025). ApJ, submitted.