whether we can change the hyper-paremeter of grpo if we donot have enough data,like batchsize ,rollout..?and i wonder the format of prompt when test on coco,whether the categroy queries are the image's gt categories or all the category of coco.thanks

[john1226966735]

[Mountchicken]

Hi @john1226966735
Q1: First, the GRPO stage does not require a large amount of data. In the original paper, we used only 100 training samples and still achieved a substantial improvement in model performance. Second, these hyperparameters can all be tuned, but I have not conducted an ablation study, so I am not sure what the optimal settings are.

Q2: When testing on COCO, for each image, we input all GT categories that are present in the current test image at once, rather than inputting all 80 categories or inputting one category at a time.