Salesforce/blip3-kale

BLIP3-KALE是一个包含2.18亿个图像-文本对的开源数据集，结合了网络规模的知识和详细的图像描述，旨在促进多模态预训练研究。数据集创建过程包括初始知识增强和扩展两个主要阶段，使用了CogVLM-17B和Mistral-7B Instruct v0.2模型。数据集可能继承自网络规模数据集的偏见，并可能存在一些管道伪影。数据集以Apache2.0许可证发布，主要用于研究目的。

BLIP3-KALE is an open-source dataset of 218 million image-text pairs, featuring knowledge-augmented dense captions combining web-scale knowledge with detailed image descriptions. The dataset creation process involved initial knowledge augmentation and scaling up stages, utilizing various models and techniques to generate and augment captions. It is primarily used for multimodal pretraining research and has been applied in multiple projects. The dataset carries certain biases and risks, such as inheriting biases from web-scale datasets and potential pipeline artifacts. The dataset is licensed under Apache 2.0, primarily for research purposes.