ContinuaL Instruction Tuning Dataset Card

Dataset details

Dataset sources

This dataset is constructed using publicly available and commonly used instruction tuning datasets, including VQAv2, VizWiz, ScienceQA, TextVQA, GQA, and OCR-VQA.

Additionally, to enhance diversity, we introduce the classification task and referring expression comprehension task into CoIN with ImageNet, RefCOCO, RefCOCO+, and RefCOCOg.

Before proceeding with instruction tuning, please download these datasets on your own.

Instruction templates

To comprehensively evaluate performance, we construct instructions using two types of templates.

License: Creative Commons Attribution 4.0 International;