Questions about the reward curve and bert.

[zjuAIHz]

Many thanks for conducting this excellent work!

I raised 2 questions while trying to reproduce the experiments.

1. Why does the reward curve for prompt-image alignment shown in this repo fluctuate greatly, but the reward curve shown in Figure 5 in the paper is very smooth?
2. While I was experimenting, the bert scores are extremely high. For example, for the noisy image below, LLaVA generated a description: "In the image, there is a colorful, abstract, and blurry background with a mix of colors and patterns. The background is filled with a variety of colors, creating a visually interesting and dynamic scene. " The bert score of this description with the prompt "a pig riding a bike" is about 0.82. I think this may be due to the fact that the bert model I used is different from yours. So I would like to know more details about the bert used in your experiments.

Very thanks if you can give me some help :-)

[mao-code]

Same question here

[Lil-Shake]

Same question here. Have you solved it? Many thanks:)