Evaluation on general MLLM datasts

[RussRobin]

Hi, thanks for this wonderful work.

I wonder if you are willing to release scores on general MLLM benckmarks like MME, MMBench, GQA... Is there any improvements by training on COST dataset and LLaVA datasets together?

Also, in COST dataset, after unziping all files, what is seg_images folder? Where do images here come from?

Thanks in advance!