Question about the training config

[deadlykitten4]

Hi, thanks for your nice work!

I notice that the default of max_completion_length=256, while I want to train on my dataset, where 1024 leads to the best results. I am not certain about the configuration. Do I need to set max_completion_length=1024 for training?

[siyan-zhao]

Hi,

Yes you can just modify https://github.com/dllm-reasoning/d1/blob/main/diffu-grpo/slurm_scripts/train.yaml#L60