stage2.sh

[lhqxcc]

Thank you very much for your outstanding work. When using your stage2.sh script for training, I encountered two questions:

1. Is the llava-158k you used composed of complex_reasoning_77k.json, conversation_58k.json, and detail_23k.json combined, or is it separately llava_instruct_150k.json?

2. In your script, you wrote: dataset_name VisualGenomeDataset@InstructCaptionDataset@VQAv2Dataset@AOKVQADataset@FlickrEntityDataset, but it results in the following error.

** module 'pink.datasets' has no attribute 'InstructCaptionDataset'. Did you mean: 'PretrainCaptionDataset'?**

Can I directly replace it with LLaVA?

[SY-Xuan]

Sorry for the typo. This dataset should be LLaVADataset. You can use the dataset of LLaVA v1.