Did anyone try to launch this repo with a more recent version of veRL?

Hi, I tried to replicate this exact repo, but with veRL version 0.90, but I couldn't get it to work no matter what I've tried. I tested many different hyperparameters and I tested both Qwen2.5-1.5B and Qwen2.5-3B and I can't get more than 15% accuracy.

Does anyone knows what might have changed in the newer versions of veRL that could make it harder to replicate TinyZero training? I know veRL applies the chat template of the model by default (which for base Qwen2.5 is the same as for the instruct version), and the older versions of veRL did not, however I disabled it.

Or maybe someone have already tried running TinyZero on more recent veRL versions and succeeded? I would appreciate any scripts or tips on reproducing this repo on newer versions of veRL.

Thanks!