This paper introduces a benchmark dataset that can be used to compare the performance of statistical postprocessing methods for several key weather variables. Such a dataset is very useful to the postprocessing community in which numerous methods have been (and continue to be) developed but are often difficult to rank because they have been tested in different setups. While particular setups often require tailored solutions, at least subgroups of postprocessing methods can be considered to have interchangeable areas of application, and a benchmark data set can thus be useful to identify methods that stand out as particularly powerful. Along with the introduction of the data set the authors use the example of 2 metre temperatures to demonstrate how such an intercomparison of methods could be performed. I welcome this initiative, and I only have one minor comment and a few suggestions regarding language.

Minor comment:

Section 4.8: Is there a reference for this ANET method? If a reference is provided, the description of the method is adequate. Otherwise, more explanation is required for details like the dynamic attention mechanism. Since methodology is not the focus of this paper, such information could be provided as supplemental material.

Typos and language:

60: complement -> complemented

113: as aforementioned -> as mentioned before

198: significant tests are run -> statistical tests are performed

216: validity -> valid

345: 'as' seems unnecessary here

This paper introduces a publicly available dataset meant to help standardize and compare studies of various methods for postprocessing ensemble forecasts.  The paper is welcome because of the substantial ongoing activity in developing such methods.  Much of this recent activity has involved machine learning methods, and I personally have noticed that very few such studies adequately (including computation of possibly statistically significant differences) compare their results to the more conventional, established methods.  The dataset described here should facilitate such comparisons, to the benefit of the discipline and quality improvements in weather forecasts.  The paper should be accepted after consideration of the following comments:

The paper would benefit from copy-editing by a native English speaker, to correct the grammatical errors that occur in various parts of the paper.  

The caption for Fig. 1 should include an explanation of the various rectangular regions.

Line 184.  Disagree that bias characterizes skill of anything.  “Skill” refers to relative accuracy, and bias is not a measure of accuracy.  For example, unbiased forecasts can result from equal numbers of extreme (and therefore poorly skilled) overforecasting and underforecasting. 

Line 186. Equality of average ensemble standard deviation and RMS error of the ensemble mean is indicative of calibration only if the forecasts are unbiased, because (squared) bias will inflate RMS error but will not affect ensemble standard deviation.  This metric therefore cannot distinguish bias from ensemble underdispersion. Rank histograms or reliability diagrams are better indicators of calibration.  To what extent do the nonzero biases in Fig 3b affect the spread-error-ratio results in Fig 3d?   Similarly, in Figure 6 (line 335ff).

Equation 1.  The overbar notation usually denotes sample average.  Perhaps use a “prime” accent instead to denote the anomalies.

line 247.  Should specify that the preserved correlations are simultaneously spatial, temporal, and intervariable.

Section 5.  It would be worthwhile to give at least an indication of the sampling variability of the presented verification statistics, if not full hypothesis test results, in order to give the reader an idea of the magnitudes of differences that are meaningful.  In particular, I am interested to see whether the substantially greater computational overhead of the machine learning methods results in significantly better performance.  Line 322 notes that “most postprocessing methods perform similarly”.

The paper introduces the first version of the EUPPBench benchmark dataset for statistical post-processing, which contains both gridded and station data of several weather quantities and different constant fields as well. The idea behind the initiative is very clear as this analysis-ready dataset provides a standardized tool for testing new approaches to post-processing and comparing their forecast skill to the existing well established methods. Besides introducing the dataset itself, the authors provide a very useful example together with the corresponding R and Python codes. They calibrate 2m temperature forecasts with the help of the state-of-the-art post-processing methods ranging from simple accounting for systematic and representativeness errors and reliability calibration to highly advanced machine learning-based distributional regression networks and ANET. The paper is well written, but far from being self-contained, one has to follow the provided references to understand completely the considered methods. However, this is not a disadvantage, as the main goal is to demonstrate the utility of the benchmark dataset. I have just a few minor remarks to be addressed.

1. P1L13: "learn about the models weather forecast deficiencies": please reformulate.
2. P2L23: and P24L503: Instead of the arXiv preprint of Lakatos et al. 2022, please cite the published one (Lakatos et al. 2023)
   
   Lakatos, M., Lerch, S., Hemri, S. and Baran, S.: Comparison of multivariate post-processing methods using global ECMWF ensemble forecasts, Quarterly Journal of the Royal Meteorological Society, https://doi.org/10.1002/qj.4436, 2023.
3. P15L93: please provide some reference to ANET or a more detailed description.
4. Following the one shot installation process of the codes suggested in https://github.com/EUPP-benchmark/ESSD-benchmark, I was not able to download code for ANET. Please check.