do you mind posting the logs and wandb results your experiment?

[yxchng]

it will be very helpful for reproduction.

[yutaocui]

Yes, we have the plan to release the training logs.

[tulvgengenr]

@yxchng Thank you for your question. We plan to release detailed experimental logs in the future. Here is a comparison of the loss curves of MixGRPO and DanceGRPO with HPSv2 as single-reward. You can use this as a reference.

[yxlabs]

thanks for the curve, just to clarify

1. is dancegrpo just trainingstrategy="all"?
2. is this curve with unified reward or without?

[tulvgengenr]

@yxlabs

1. Yes, in addition to setting training_strateg to “all,” you also need to comment out --flow_grpo_sampling and enable --drop_last_sample. This is because the flow-GRPO and DanceGRPO sampling strategies have slight differences in their formulas. Finally, set eta to 0.3, which is the default setting for DanceGRPO.
2. No, this is only with HPSv2 as single reward.