Very good question! These models are trained on ~40 years of ERA5 data –you can think of past forecasts from numerical models integrated with real observational data to have a continuous distribution of weather parameters (temperature, wind etc.) Therefore model resolution is 0.25 degrees (28km x 28km at the equator).

The way accuracy is measured is through picking targets (say temperature at 2 meters, at x,y lat & lon and forecasted 24h ago) and comparing them on RMSE and ACC (anomaly correlation co-efficient). For instance, in Google Graphcast paper they pick 1380 targets and the model out performs NWPs in 90% of them.