A valuable dataset is presented in this manuscript, and it also addresses an important problem in satellite lightning geolocation. My concerns are regarding the corrections that are dependent on reference datasets. In addition spatial correctness, and validation strategy need to be addressed before the dataset can be considered fully reliable for broad scientific applications.

1. Physics-based corrections (e.g., payload misalignment, cloud-top height parallax, thermal deformation), followed by an additional empirical correction based on the WWLN were applied. This combination is reasonable, but it remains unclear why a substantial residual empirical correction is still required after applying physically based models. Whether these arise from CTH uncertainties, instrument/navigation limitations, or incomplete physical representation.
2. Authors produced a “unified correction dataset” using subregional (20 × 20) curve fitting. But that depends on lightning occurrence. This questions about the spatial consistency, especially in data-sparse regions.
3. The manuscript relies on WWLLN. It has biases due to its low and spatially variable detection efficiency, bias toward strong CG strokes, and location uncertainty (several km). The authors should therefore: Quantify residual error after physics-based correction alone, Justify the need for the empirical step, and Discuss how WWLLN limitations affect the robustness of the final dataset. Authors partially justified the use of the WWLN (by its global coverage); however, this should be balanced against its known limitations in detection efficiency and accuracy.
4. Recent studies (e.g., Wu et al., 2024, https://doi.org/10.1016/j.jastp.2024.106194) using higher-quality regional networks, such as the National Lightning Monitoring Network, have shown significant discrepancies in FY-4A LMI performance. The authors should:

• Discuss the trade-off between coverage and data quality,
• Evaluate how WWLLN limitations affect correction accuracy, and
• Consider (or discuss) the role of regional high-DE networks for validation or hybrid approaches. The exclusion of oceanic regions further indicates that the correction is not uniformly applicable, which should be explicitly acknowledged.

1. The authors should include case studies comparing lightning locations before and after correction with CTT/CTH and/or radar reflectivity, to demonstrate improved alignment with convective structures. Where possible, comparison with regional networks such as the China Lightning Detection Network or National Lightning Monitoring Network would strengthen validation.
2. Please add a discussion of remaining limitations after correction is needed, including residual errors (~15 km), dependence on WWLLN, and implications for different applications (storm-scale vs climatology). This is particularly important given reduced reliability in low-lightning regions, as acknowledged by the authors.
3. It is not clear whether the correction varies with viewing geometry (e.g., satellite zenith angle) or latitude; this should be discussed.

Minor Comments

1. The description of the matching criteria between FY-4A LMI events and the WWLN should be clarified, specifically the choice of spatial and temporal thresholds.
2. Please clearly explain how the 20 × 20 subregional division was selected. Are the results sensitive to grid size?
3. The authors should explicitly state whether day–night differences in LMI detection efficiency were considered in the correction process.
4. Uncertainty estimates associated with the fitted correction functions are not clearly presented and should be included.
5. The reported ~15 km residual error should be discussed in terms of its implications for different applications (e.g., storm-scale vs climatological studies).
6. The manuscript would benefit from a brief comparison with other space-based lightning sensors, such as the GOES-R Geostationary Lightning Mapper or TRMM Lightning Imaging Sensor, to place the results in context.
7. The terminology distinguishing event, group, and flash should be defined clearly at the beginning for consistency.

The authors describe a geolocation correction methodology for the FY-4A lightning mapping imager (LMI). They compare to ground-based lightning detection (WWLLN) in post-processing to produce a uniformly QC’d dataset for the full period of operation of the satellite.

The manuscript is clear and the availability of the data and its characteristics is most welcome. 

I have a few minor comments below, which are all easily corrected.

Line 49: It will be possible to apply the authors’ correction method to FY-4C LMI data, but not in real-time since it relies on other lightning data. Is such a post-processed dataset planned? More generally, will lessons learned from FY-4A mitigate the need for such a method due to improved operational algorithms for FY-4C?

Line 75: How are thermal distortions  (forced by the diurnal cycle) related to launch-induced payload displacement? Presumably those would occur even in the absence of any launch effects.

Line 129: what were the other “ground-based lightning observations” used by Fan et al. (2018) for comparison with WWLLN?

Line 131: “of [WWLLN is] around 10 km”

Line 141: please provide a reference for “marginal effect analysis”

Line 179: “the weight assigned to the 𝑘-th data point” is defined as omega sub k with an overbar, but the formula also has a tilde above the overbar. In the definition of Eq. 2, there is an additional notation that adds a superscript “raw” to omega. Please clarify how each of these symbols are related to one another, and their possibly varied role in the spline fitting process. For example, is Eq. 2 only a starting estimate for omega?

Section 3.1: matching of WWLLN strokes to LMI events is described here. I would have expected matching to group centroids. In the case of the a bright optical pulse comprised of many pixels, does the authors’ fitting procedure only include those events that are in the 3s and 30 km space/time window? Are the events outside this window simply discarded in the fitting analysis?

Please mark the four subregions shown in Fig. 4 on Fig. 3

Line 243: I don’t see an east-to-west propagation of the error pattern. To me it looks more like a steady amplification of regions that always have a relative maximum of errors.

Line 289: Typically, more samples should lead to improved location, but the conclusion here is that a low number of samples leads to better fits. Why? Does this suggest a shortcoming in some part of the fitting approach?

Line 291: I’m not convinced that the large errors in the NE are due to the large ground footprint of the detector elements. If that were the case, the NW part of the domain should exhibit the same pattern.

Line 300: While I don’t object to distribution of a CSV file format, which is easily understood and machine readable (if lacking in terms of metadata), it is not true that NetCDF files will be larger. Naively reading the 20190406.csv dataset (49.2 MB) with pandas and writing it back to NetCDF with xarray was a bit smaller (48.7 MB), even naively using 64 bit ints and floats for all columns. However, the file can easily be made much smaller. Year, month, day, hour, minute and second can all be stored in much fewer than 64 bits, and packing the floats as either 16 or 32 bit floats, or as ints using scale_factor and add_offset, will save further space. Finally, internal zlib compression can also be used on NetCDF files and will be transparently decoded by the library. Finally, the authors might also consider using the CF metadata standard to encode time as a single variable.