TaskGalaxy
收藏资源简介:
TaskGalaxy数据集是由快手科技提出的一个大规模多模态指令微调数据集,包含19227个层次化的任务类型和约413648个视觉问答样本。该数据集通过利用GPT-4o从少量手动定义的任务类型出发,自动扩展出多样化的任务类型,并通过CLIP模型和GPT-4o生成相关的问题答案对,再通过多个模型筛选以保证数据质量。该数据集在多模态场景中极大地提升了任务类型的多样性,可应用于提升多模态模型在各类任务中的泛化能力。
TaskGalaxy Dataset is a large-scale multimodal instruction tuning dataset proposed by Kuaishou Technology. It encompasses 19227 hierarchical task categories and approximately 413,648 visual question answering (VQA) samples. To construct this dataset, GPT-4o is first employed to automatically expand diverse task categories from a small number of manually defined task types. Then, relevant question-answer pairs are generated via CLIP and GPT-4o, followed by data filtering with multiple models to ensure data quality. This dataset greatly enhances the diversity of task categories in multimodal scenarios, and can be applied to improve the generalization ability of multimodal models across various tasks.




