Question about model training reproduction

[HzRynn]

I’m reproducing the LTF model with default and simulation co-train settings, but my EPDMS results are 1~3 lower than reported in the paper / README.
Are there any extra implementation details I missed for the training setup?

Thanks!