Scores are always negative when using EditRewardInferencer with reward_dim="overall_detail"

[tls0523]

Hi, thank you for releasing this great project 🙏
I’m trying to use EditRewardInferencer for single-image inference, but I noticed that the returned reward scores are always negative, and I’m not sure whether this is expected behavior or a potential issue in my usage.

My setup
I initialize the inferencer as follows:
def __init__( self, device="cuda", config_path="EditReward-Qwen2.5-7B-VL.yaml", checkpoint_path="EditReward-Qwen2.5-VL-7B", reward_dim="overall_detail", rm_head_type="ranknet_multi_head", ): self.inferencer = EditRewardInferencer( config_path=config_path, checkpoint_path=checkpoint_path, device=device, reward_dim=reward_dim, rm_head_type=rm_head_type, )

During inference, I obtain the score via:
score = rewards[0][0].item()

Issue
No matter what inputs I use, the resulting score is always negative.

Questions
Is it expected that the output reward score for reward_dim="overall_detail" is always negative?