This manuscript describes a newly created dataset comprising broadband seismic recordings of earthquakes and ambient noise in Italy, called INSTANCE. Public datasets are of utmost importance to advance Machine Learning in seismology, and ensure their reproducibility. I therefore applaud the authors for their efforts to create a dataset with Machine Learning applications in mind. The authors perform various quality checks and provide metadata based on which the user can determine which data to include in their investigations. Moreover, the INSTANCE dataset could be suitable for analyses outside of the domain of Machine Learning, and the authors facilitate these by including detailed metadata of both the earthquake and noise recordings.

The manuscript that describes the dataset is well written and provides an extensive analysis of various summary statistics of the dataset, as well as a useful review of other datasets that are currently available. I have a few minor suggestions for clarifications in the text (see below), but other than that I think that the manuscript is in good shape.

The project webpage and GitHub repository are clearly structured, and include Python notebooks to reproduce the figures presented in the manuscript. The example data represent only a subset of the data analysed in the manuscript, and so the figures look different, but I understand that it could be too computationally expensive to re-run the analyses on the full dataset. In principle the users could apply the same code to the full dataset if they wish to exactly reproduce this study. One suggestion I would like to make here is to consider some form of versioning of the INSTANCE dataset. It is not unimaginable that the dataset be modified or expanded in the future, for instance to re-classify misclassified earthquake or noise traces, or to include future seismic events of great significance. By including a version tag (and changelog), users can specify which version of INSTANCE they used in their analyses, which would improve the reproducibility of future studies.

As for the core data files (I reviewed the sample dataset): the waveforms are stored in datasets labelled in correspondence with their trace name, which makes it convenient to extract only the waveforms that match a metadata query. I did notice that in many cases (about 1 in 3 traces), the evaluation of EQTransformer has yielded only NaNs. I’m not sure how to interpret this, because I would expect an integer number (zero if no detection was made) instead of NaN. Perhaps the authors could check this and mention how NaN should be interpreted, or replace them with zeroes where applicable. Nonetheless, the fact that EQTransformer fails in 1 out of 3 cases is a bit startling, and underscores the need for dedicated datasets for specific regions to (re)train Machine Learning models. Lastly, I noticed that after bzip decompression the data files take up a lot more disk space, so it could be helpful for users to indicate on the INSTANCE webpage what the data size after decompression is.

Overall, I think that INSTANCE is an important contribution to the seismic community, and I recommend acceptance of the manuscript after some (very) minor revisions.

Kind regards,

Martijn van den Ende

Minor comments on the manuscript (P = page; L = line):

1.    P1, L19-20: in addition to the URLs, I would add the names of the ML platforms to make these more recognisable (and in case the URLs are changed in the future).

2.    P6, L1-2: I don’t quite understand that is meant here. Could the authors clarify this?

3.    Figure 3, panel d: I would plot these data on a linear scale to better see the trends. It doesn’t seem very logical to me to plot azimuth on a log scale.

4.    Figure 1, panel a: the colour scale does not really provide any insights into the depth of the vast majority of earthquakes on this map (practically all the events are coloured deep red). To make this panel more informative, I would tailor the colour scale to the crustal earthquakes and accept clipping for the deep earthquakes. Also, the jet colourmap is hard to interpret in grayscale (when printing) and by readers with colour vision deficiencies. I would recommend re-rendering this figure with a perceptually uniform colourmap like viridis or cividis.

5.    P16, L8, Fig. 6: maybe I missed a mention of this earlier in the manuscript, but which velocity model is used to calculate the theoretical arrivals / residuals? Is the model consistent for the entire dataset?

6.    P17, L11-12: since the mean is removed from each trace, I would be surprised if the mean would be anything else than zero. The figures that show histograms of the mean are therefore not very informative and could be omitted, in my opinion.

7.    P27, L11: could the authors confirm that the 120s time window is different for each station? In the way it is written now, it is not fully clear to me. I think it’s unlikely that one would find a 120s window common to all stations that matches all of the criteria simultaneously, but this is not explicitly mentioned in the text. This would be an important consideration when making any assumptions regarding the spatial coherence of the noise.

8.    P32, L4: personally I would also have created a dataset with low SNR waveforms to complete the “spectrum” from noise to clear earthquakes (useful for microseismicity studies), but I understand that this may be beyond the scope of the project. Perhaps this is something that the authors could consider including in the future (see my comment about versioning).

Review by John Clinton

This is an excellent article introducing a new dataset targeting machine learning applications using a near-complete set of earthquake records from Italy. It follows the example of STEAD described in Mousavi et al, 2019. The manuscript provides an overview of how the dataset was generated and organised, and then provides an overview of the general features of the dataset.  The manuscript is well written, particularly the introduction that gives a strong motivation for why this type of product is sorely needed, and provides an overview of similar datasets. It is timely and important that high quality local earthquake records datasets outside the US W. Coast are highlighted and made easily available for researchers to use. I hope this sort of documentation of datasets and their preparation in research-ready format becomes the standard, and I expect that publication of this manuscript will lead to numerous publications on ML that use this dataset.

I know some information on the uniqueness of the dataset is dispersed through the article, but I suggest, in a single place,  in the discussion, the authors extend and accentuate what is different with this collection compared to others, besides from the obvious that this is solely an Italian dataset based on the Italian earthquake catalogue. Are the metadata fields better? How do they differ? Are formats modified ?  Is it unique to provide both raw and corrected waveform data? Data volumes similar to other datasets? Maybe a comparison table to STEAD and the Caltech datasets would be helpful. The authors could also provide stronger comments on the benefits of standardisation of formats / metadata for these datasets.

A comment on the metadata on earthquake parameters summarised in Table 2: numerous fields are provided, including location uncertainty, but in the text, there is no comment on what location algorithm or velocity model is used. Since the dataset spans 15 years and a very wide geographic range, its likely that despite efforts to ensure a continuous approach to manual catalogue review, the velocity model and location algorithms have changed across the catalogue. If they have not changed, this point should be made. If they have changed, this should be indicated in the document, and consideration should be given to add this information in future updates of the dataset.

The DOI of each of the network codes used in the dataset / publication must be made available in the references or in the data availability section

In general, the figures are not optimal, often use strange axis labelling (that may be a direct metadata field from Table 2 - if so mention it!), and often have captions that are too terse / insufficiently descriptive. I suggest the authors look through these carefully. Also font sizes on  Figs 14,15,16 are too small- in particular the exponents are completely illegible. 

Generally, the standard of English is very high. In the following are direct places in the text where I suggest technical clarification or language improvements:

p1 l17 dispose of-> make available

P2 l6 …Ml into seismology has shown the… -> …Ml in the field of seismology has highlighted the…

P2 l28-30 rephrase sentence 

P3 l 14 good -> impressive ; are -> have been

P4 l18 I know it is clarified later - much later - but maybe add here that the preferred INGV catalogue magnitude is used here, mainly Ml but sometimes also Mw and Md

P4 l21 taking -> using

P4 l25: some >M4.0 are rejected. A bit more info on significant events that have been removed is needed. Are these only those that include multiple events in the same time window in the catalogue? I hope no very significant events are rejected simply because a very small foreshock or aftershock is also catalogued…

P4 station selection: mention in this section that only stations on Italian territory are used. Are the Civil Defence stations not added? If not, mention why this very significant dataset missing - is it technical or political? 

P5 l1-5: Be more specific on what picks are made available. I assume the INGV catalogue makes first arriving P and S picks only. No additional phase type is indicated (Pg vs Pn), and secondary phases are also not identified (eg PmP)

P5 l6-9 its should be accentuated either here or later that since 1/ not all stations used in the catalogue generation are included, eg foreign-operated stations; 2/  phases with large residuals or low weight are removed that it is not possible to use this dataset to relocate the catalogue.

P5 waveform data selection: in the case of multiple available sampling rates, I infer that the same sampling rate used to make the manual pick is selected for inclusion here. Or is it the highest available sampling rate for each channel?

P6 l 7 starting -> start

P6l11: I don’t understand what is ‘arrival time samples’. Why not simply use time in seconds?

P6 l 19 Title does not read well. ‘Data counts waveforms’ is not clear. I suggest rephrase: ‘preparation of processed waveforms in digital units’

P6 2.1.5: the authors should mention are all traces rotated into ZNE, or in the entire Italian catalogue in ZNE by default. If so, I am amazed!

P7 l16 … generated the associated ground motion units dataset after… -> … generated a dataset in units of physical ground motion after …

P7 l 23 result -> be

P7 l.29 2.2 Metadata: in source, the location method or velocity model are not included. They should be if either of these have changed over 15 years of the catalogue.

P8 l25 ‘missing data’ - please expand

P8 l 28 onset -> onsets

P8 l 31 upon -> using

P9 l 6 besides -> in addition to the fact

P9 l11 analogously -> similarly

P9 l12 remove ‘found’

P10 Table 2: location code is not part of the International Registry?

P12 l 4: outside -> in the near vicinity to

P12 l9/10 check red / blue for normal / inverse - seems these are switched in the figure / text. Use either thrust or inverse in both text / figure.

P12 l14 evidences that quite different…-> demonstrates that quite a different….

P13 l3: …it difficult the phase picking -> the phase picking difficult

P13 l 5 I’m surprised to see selection criteria was for even number of traces for each channels? Seems in contradiction to 2.1.2, where all reasonable phases according to seismicity were selected. Was seismicity for smaller events actually selected according to numbers of station pick?

P13 l6,7 - revise sentence, clumsy.

P14 fig 4 caption: diagrams -> diagram

P14 l3 remove ‘included’

P14 l4 ‘great majority’ seems an exaggeration.

P15 Fig 5 / l9 onwards: the number of up first motion polarities is double that of down. This is surprising, and possible concerning unless there is a reasonable explanation I do not see. The authors should explain this. Is it possible eventype=earthquake is not selected, and blasts are also included here?  

P16 l10 wide range ‘of’ waveform…

P17 l1 1/2: is it possible this can also be explained by systematically mis-identified first arrivals, rather than complications in the velocity structure?

P17 l 12 evidence -> display or feature?

P18 l9/10 rephrase sentence!

P18 l11 rather satisfying -> sensible?

P19 l1 at higher values -> at higher and lower values

P19 l4 …IMs can be assimilated to an -> IMs represent an

P20 Fig 10: add to caption that top shows linear y-axes, bottom shows logarithmic y-axes.

P21 l7: ‘exhaustive’ is quite an exaggeration - its more to show how metadata can be used to isolate end members.

P24 Fig 14: would be good to mention here and elsewhere that the fields in italics can be defined in Table 2.  (Same for Fig 15, 16)

P27 l10: earthquake in INGV catalogue - so its very possible that noise traces include energy from regional and telesiesmic events.

P27 l14: any effort to include the same spread of stations as found in the event dataset?

P27 l17 46 metadata -> 46 metadata elements

P33 l 16: reveal -> prove

P34 l 35 Sesbench -> SeisBench

P39 FigA4 - over 100 records have PGA >2g, and many even over 4g. Which is rather unphysical.  Is this understood?