wandb and difficult to reproduce

[lujiarui-iie]

Thanks for your fantastic work!

It seems that Wandb only provides training information for SFT and does not provide relevant training information for PPO. Is the provided Wandb link the training process of PPO? Or is the PPO training process after SFT?

It is difficult to reproduce PPO results using the default parameters in train_ppo.sh and the same environment, and even training cannot converge.

Can you give me some advice?

[quao627]

Thanks for notifying us! We are working to reproduce the issue and ensure that the GitHub code produces the expected results. Please stay tuned. We will address this as soon as possible.

[jiangxinke]

Hi there!
I wonder if you could share the Weights & Biases (wandb) link for your trained model? My reward curve is significantly lower than the results reported in your paper—even lower than Search-R1—and I’d love to cross-check configurations or training details to identify potential issues.

Thanks a lot for your help!