Difficulty reproducing forecasting results with identical config and pretrained weights

[Gwenielulu]

Hi, I followed the official config (configs/global_forecast_climax.yaml) and used exactly the same hyperparameters as described in the paper. All training parameters — including learning rate, optimizer settings, and warmup epochs (set following the official code) are identical to the original setup. The pretrained checkpoint (5.625deg.ckpt) was loaded successfully, but the forecasting results, especially for geopotential_500, still deviate significantly from the reported values (≈3× higher RMSE), while other variables remain consistent or slightly better.Could you please confirm if there are any hidden preprocessing, normalization, or climatology adjustments affecting geopotential metrics?

My experiments are based on the WeatherBench2 ERA5 dataset, which may differ slightly from the original preprocessing used in the ClimaX paper. I wonder if this could partially explain the higher RMSE observed for geopotential fields.

Thanks for your great work and open-sourcing ClimaX!