cota-mantis
收藏资源简介:
TACO数据集包含293K条由GPT-4o生成的合成思维链和行动链,涉及15种行动,如OCR、深度估计、计算器等。这些数据用于微调多模态语言模型,以生成思维链和行动链来回答复杂的视觉问题。数据集的生成过程包括使用GPT-4o和Python程序生成大量的合成数据,并通过数据过滤和混合技术获得高质量的子集。数据集的来源包括Cauldron和Mantis-Instruct,这些数据集收集自COCO、AOKVQA、ScienceQA、Visual Genome等多个现有数据集。
The TACO dataset contains 293K synthetic chain-of-thought and chain-of-action samples generated by GPT-4o, covering 15 categories of actions such as OCR, depth estimation, calculator utilities, and others. This dataset is intended for fine-tuning multimodal large language models to generate chain-of-thought and chain-of-action sequences for answering complex visual questions. The generation pipeline of the dataset involves generating large-scale synthetic data using GPT-4o and Python scripts, followed by data filtering and mixing techniques to obtain high-quality subsets. The TACO dataset draws its sources from Cauldron and Mantis-Instruct, which are compiled from multiple existing datasets including COCO, AOKVQA, ScienceQA, Visual Genome, and more.




