Yichuan

Oh, I see — that’s very clear now!
By the way, I still have one question: maybe it would be better to make the evaluation easier to score, for example using exact match or F1.
I was also curious how well an LLM alone (without the images) could handle these queries — I tried a few, and for some questions, it actually did quite well.

Hi, QQ here. How do you measure the sownstream generation quality here, since you have a golden label like answer in the query.
If I want to measure the generation quality, I am not sure what I can do, I don't quite understand how you label the answer, though