How are QeRL-trained models evaluated on other benchmarks?

[LJlimo]

Hello, and thanks for your great work.

The repo only released evaluation scripts for the base model without QeRL training on MATH500, AIME24, and AIME25. I'm just wondering how to evaluate QeRL-trained models with LoRA checkpoint on these benchmarks.

Any advice would be greatly appreciated.
Thank you for your help.