General comment

This paper compares a stand-alone LSTM model with a hybrid HBV-LSTM model on the CAMEL-CH dataset. It also examines the impact of training the stand-alone LSTM on the CAMEL-US dataset, alongside CAMEL-CH, and also using 256 nodes instead of 64. The main focus of the discussion is on the ability of both models (stand-alone LSTM and hybrid) to show a linear pattern between simulated peak flows and rainfall when applying "synthetic" rainfall far higher the observed ones. The results clearly show the impossibility for the stand-alone LSTM model to exibit this linear pattern due to simulated discharges tending to a limit values as predicted by a previous study. But more interestingly, it clearly shows that this observed limit is far lower than the theoretical limit expected by the authors.

In my opinion, this paper is very interesting as it is the first to clearly and honestly address the limitations of LSTM models in hydrology. 

I recommend publication after revision.

Main comments 

My main comments are:

1) Even though it is outside the scope of the paper, I would have appreciated a "deeper" and "fairer" comparison between the stand-alone LSTM model and the hybrid HBV-LSTM model. The paper is short (only 3 figures of results in the main text), there is room for that. My main criticism is that no hyperparameter (HP) tuning is done for either model. The HP values are simply taken from previous studies. I think that the results could be different if a proper fine tuning was done for each model.

2) For analysis, it would also be very interesting to see the results for a single HBV model as a benchmark, which is very "cheap" to calibrate locally. Is there an improvement and is it "worth" the huge amount of data and GPU time required to process it? For example, the authors added US CAMEL data to their CH CAMEL learning dataset and moved from 64 to 256 nodes, which would have required a considerable amount of additional resources, but they don't show the corresponding improvement.

3) As the paper focuses on extremes, I also think that the evaluation against the observed runoff should not be limited to the NSE criteria as in Fig. 1 (which is the only figure presenting models performances), but should include a deeper analysis, including for example signatures calculated on flood events.

4) The same comment applies to the second part of results (Figs. 2 and 3, using synthetic rainfall): only 1 flood for 3 catchments (a little more in the appendix), whereas the authors have thousands of examples. A synthetic metric should be found that "summarises" the different observed behaviours (between catchments, but also for the same catchment but under different conditions). A "visual" analysis on a few examples, as in this paper, is a first step to draw first hypotheses. But then these hypotheses should be tested in depth..

5) This last point (the need for a synthetic metrics that allows a "deep" analysis) leads me to my main comment. The authors don't clearly explain why, from a hydrological point of view, peak discharge should increase linearly with extreme rainfall. I fully agree with this, and even if it seems obvious, I think it would be valuable to anchor the paper with more basic hydrological references. In terms of synthetic metrics, I would, for example, calculate a regression coefficient between peak discharge and synthetic rainfall and see how it changes as a function of rainfall, as in the paper, but also as a function of the initial moisture content before a flood and/or the runoff coefficient. I would also not look at flood by flood, but try to find a graphical representation of all floods and catchments together.

Minor comments

L100 : why did not you do a hyperparameter tuning? (a LSTM expert told me one day that  hyperparameter training is absolutely required in any case, and that, if "hydrologists" don't have the necessary GPU resource, they should not use LSTM)

L200: You should give more details on the models performances, for instance using flood signatures

L224: The results for the 256 node LSTM and/or the training using US-CAMEL should be presented in Fig .1 and discussed. Does this huge amount of additional data improve models performances?

L235: You should do more clearly the link with basic hydrological processes, such as soil saturation and the effect of initial humidity condition.

Figure 2 and 3 : the terme "observation" is misleading. There is no observed discharges in this figure.

L260: this affirmation is supported only by 1 flood over 3 cathments. You should try to exhibit that using much more discharge simulation (...that you have)

L299 : "Extreme hydrological events often coincide with distinct regime shifts": I fully agree but could you explain what do you mean to a "non-hydrogist", in term of involved processes.