Training is much slower than reported and configs seem inconsistent

[lingwei-zhang818]

Hi, thanks for releasing this work.

I am trying to reproduce the MediX-R1-8B result using training/examples/medix-r1_8b_dapo.sh on 8 H100 GPUs. I basically followed the released script, but the training speed looks very different from what is reported in the paper.

In my run, the training took about 107 hours, while the paper reports around 25 hours. I am not sure if I missed some important setting, or if the released script is not exactly the same as the one used for the paper.

I also noticed that some hyperparameters in medix-r1_8b_dapo.sh seem different from Appendix A.2. Could you please clarify which config was actually used for the reported MediX-R1-8B DAPO result?

Specifically, I want to confirm:

1. Is training/examples/medix-r1_8b_dapo.sh the exact script used for the paper result?
2. If not, could you share the correct hyperparameters/config?
3. For 8 H100 GPUs, is ~107 hours expected, or should it be closer to the reported 25 hours?

Thanks!