Ground-motion models (GMMs) are often used to predict the random distribution of spectral accelerations (SAs) at a site due to an earthquake at a distance. In probabilistic seismic hazard and risk assessment, large earthquakes occurring close to a site are considered as critical scenarios. GMMs are expected to perform well for such rare scenarios i.e., to predict realistic SAs with low prediction uncertainty. However, the datasets used to regress GMMs are usually deficient of data from rare/critical scenarios. The Kotha et al. (2020) GMM developed from the Engineering Strong Motion (ESM) dataset was found to predict decreasing short-period SAs with increasing \({M}_{W}\ge {M}_{h}=6.2\), and with large within-model uncertainty at near-source distances \({R}_{JB}\le 30km\). In this study, we analysed and updated the parametrisation of the GMM based on non-parametric and parametric analyses of ESM and the NEar Source Strong motion (NESS) datasets. By reducing \({M}_{h}\) to 5.7, we could rectify the \({M}_{W}\) scaling issue, while also reducing the within-model uncertainty on predictions at \({M}_{W}\ge 6.2\). We then evaluated the updated GMM against NESS data, and found that the SAs from a few large, thrust-faulting events in California, New Zealand, Japan, and Mexico are significantly higher than GMM median predictions. However, near-source recordings of these events were mostly made on soft-soil geology and contain anisotropic pulse-like effects. A more thorough non-ergodic treatment of NESS was not possible because most sites sampled unique events in very diverse tectonic environments. Therefore, for now, we provide an updated set of GMM coefficients, within-model uncertainty, and heteroskedastic variance models.