Difference between two datasets (VisionThink-General-Train and VisionThink-Smart-Train)

[LinZichuan]

Hi @Yangsenqiao ,

I noticed that in the Data Preparation section of your paper, you mentioned a general VQA dataset, but there are two datasets in the repository (VisionThink-General-Train and VisionThink-Smart-Train).

Could you explain the difference between these two datasets? Thanks!

[Yangsenqiao]

Hi @LinZichuan 👋,

Thanks for your interest in VisionThink! Both VisionThink-General-Train and VisionThink-Smart-Train are general VQA datasets, but they are used to train different models:

1. VisionThink-General-Train is a larger dataset used to train Senqiao/VisionThink-General. This model is trained on full-resolution images and showcases how LLM-as-Judge enhances general reasoning performance.

2. VisionThink-Smart-Train is a smaller dataset used to train Senqiao/VisionThink-Efficient. It adopts a two-stage process: the model first receives a downsampled image; if it determines the information is insufficient, it automatically requests the high-resolution version.

Hope this clarifies the distinction. Thanks again for your interest in VisionThink!