The manuscript “GEMS-GER: A Machine Learning Benchmark Dataset of Long-Term Groundwater Levels in Germany with Meteorological Forcings and Site-Specific Environmental Features” presents a valuable new groundwater dataset for Germany that consolidates scattered information into a unified and accessible format. The dataset includes time series of groundwater levels, meteorological forcings, and a range of static environmental descriptors. The authors have clearly invested considerable effort in collecting, harmonizing, and cleaning these data, and the result is a resource of great potential value for the hydrological and environmental science communities. Overall, I find this paper well-prepared and the dataset highly relevant. It is very positive to note that the documentation and data availability is clear and everything is easy to follow. This is a great example of how things should ideally always be. However, I believe some revisions are necessary to further strengthen the manuscript. My comments are as follows:

General Comments

Relevance Beyond Machine Learning

While the dataset is positioned primarily as a resource for machine learning (ML) approaches, its value extends well beyond this scope. The authors could broaden the framing of the paper to highlight its usefulness for a wider range of hydrological and environmental applications, including traditional modeling approaches, process understanding and decision-support tools.

Discussion of Uncertainties

Although the authors have clearly put effort into filtering implausible values and improving data quality, the manuscript does not adequately address uncertainties. In particular: i) Measurement uncertainties in groundwater levels are not discussed and ii) Uncertainties in derived or static attributes (e.g., recharge estimates, soil data) are likely more impactful for modeling and deserve attention. A discussion of these uncertainties would help readers better understand the dataset’s limitations and appropriate use.

Interpolated vs. Raw Data

The manuscript mentions interpolation of missing values in groundwater time series. While the interpolation is reasonable, every method carries assumptions and trade-offs. I recommend that both versions—interpolated and raw (with gaps)—be made available. This would increase transparency and provide flexibility for future users.

The interpretation of NSE values is debatable. Stating that NSE ≈ 0.5 is “relatively strong” is not convincing in this context, as groundwater levels—compared to surface water flows—are typically smoother and have longer response times and are typically easier to predict. NSE values <0.5, especially at a weekly resolution, suggest that model performance is modest and should be described as such.

In this context, (for example Line 264) the benchmark modeling is intended to provide transparency rather than achieve optimal predictive performance. However, the interpretation of model performance (e.g., identifying locations where other drivers may be important) seems somewhat optimistic. For instance, Figure 5 demonstrates that even at sites without apparent additional drivers, machine learning performance can still be limited. A more nuanced discussion of these limitations would strengthen the paper. Moreover, when the model performance is not fully mature, one might question the rationale for conducting this analysis, as it could potentially create confusion—particularly when the boundary between poor model performance and the influence of external drivers becomes blurred.

Additionally, it is not clear to me whether the NSE is based only on observed measurements or if it also includes interpolated values. The comparison should only be made using the actual measurements, not values that are already derived from a method or model.

Furthermore, I am wondering why no values are provided at least for the major rivers. Especially when using machine learning approaches, it is crucial to integrate the relevant processes—such as groundwater or surface water interactions—otherwise a good fit may be achieved for the wrong reasons. In my opinion, this should definitely be done, particularly because river data are generally reliable and easily accessible.

Specific Comments

Line 24: The statement that ML approaches “can/should be used for assessing the impact of climate change” is too strong. ML models trained on historical data may not extrapolate reliably to unprecedented climate conditions (e.g., prolonged droughts, multi-year droughts which may create a system trigger point). The statement should be rephrased to reflect these limitations.

Line 31: There is a modeling approach between ML and fully physically based models—simplified point-scale or lumped-parameter models (e.g., AquiMod, Pastas). As noted by Bakker & Schaars (2019;  10.1111/gwat.12927), these models can provide efficient and accurate groundwater forecasts with lower data requirements. Including this perspective would improve the model concepts completeness although the approach is not used here (https://doi.org/10.1016/j.envsoft.2014.06.003; https://doi.org/10.1016/j.jhydrol.2023.130120; https://doi.org/10.1111/gwat.12819 among other references) .

Line 34: The statement “…even when observational data are limited” is oversimplified. It should clarify that transfer learning or cross-site modeling may help in data-scarce regions, but the limitations of sparse observations must be acknowledged.

Figure 2: Please clarify why no data remain for HB after filtering. Unlike SL or HH, where the absence of data is understandable, it appears that HB has usable data that could potentially be gap-filled. More details and explainations would help.

Figure 4: This figure illustrates the type of uncertainty that deserves more attention (as mentioned above). For example, recharge estimates often vary widely depending on methodology (see for example https://hess.copernicus.org/articles/25/787/2021/, https://doi.org/10.1029/2022GL099010 although the scale is different), and this variability may affect ML performance. I am not proposing to do the modelling with multiple recharge or precipitation products (or any other variable) because this is just unrealistic but discussing these uncertainties explicitly would be helpful.

Review manuscript essd-2025-321.

The manuscript describes a important  groundwater level dataset for benchmarking machine learning models. I like the dataset because it is both important and timely for multiple disciplines, including hydrology, hydrogeology, and data science, etc. There is no doubt that it will be widely used for machine learning development, similar to the CAMELS datasets. I also like the manuscript because it is clear and concise. The dataset is well organized and follows the FAIR principles. I have few comments on the manuscript and dataset for the authors’ consideration when they do some minor revisions.

Major comments.

1. Are the groundwater levels from shallow wells, deep wells, or a combination of both? Please clarify this information in the manuscript.
2. Although the imputation rate in the dataset is quite low, would it be good to include a synthetic test to evaluate the quality of the imputation (i.e., how well or poorly it performs)?
3. The role of static covariates in machine learning model development (e.g., CNN or LSTM) is somewhat limited if data from all 3,207 wells are used for training, validation, and testing. I suggest the authors perform a spatiotemporal testing experiment — for example, hold out a group of wells entirely from the training set and use them only for testing — to evaluate whether the static covariates improve model performance.

Minor comments.

Lines 91–101: Consider presenting the names in a table and including the number of wells associated with each. This would make it easier for readers to understand.

Line 111: The acronym LANUV has already been introduced earlier; the full name is not needed here.

Section 2.2.4: Please add one or two sentences explaining the possible reasons for the identified outliers.

Line 186. You mean ‘density’ here, is that calculated by number of wells divided by area? or MHD1 has the highest number of wells?

Lines 190–200: Are these variables listed in Table 2? If so, note that they could also serve as static covariates for machine learning model development.

Section 3: While the cleaned and quality-controlled groundwater-level time series are very useful, have the authors considered including the original (raw) dataset as well?