The CKDMIP datasets are described in detail by Hogan and Matricardi (2020) available on the CKDMIP home page.
The two main datasets are Evaluation-1 (for training and evaluation) and Evaluation-2 (for independent evaluation). Each contains 50 profiles from the much larger NWP-SAF dataset of Eresmaa and McNally (2014). Seventeen of the 50 profiles are chosen because they have the maximum or minimum temperature, humidity or ozone at a particular pressure level of all the NWP-SAF dataset. The remaining 33 profiles are chosen randomly.
Figure 1. The temperature, humidity and ozone concentrations for the Evaluation-1 dataset.
Figure 2. The temperature, humidity and ozone concentrations for the Evaluation-2 dataset.