andropar/relaion2b-natural-embeddings

ReLAION-2B自然嵌入数据集包含约5亿张自然照片的预计算CLIP ViT-H/14嵌入。这些嵌入是从ReLAION-2B数据集中提取的，每张照片都有一个自然分数（0.7-1.0），用于表示照片的自然程度。数据集结构包括图像URL、自然分数、原始LAION-2B嵌入中的行索引和768维的CLIP嵌入向量。数据集主要用于图像相似性搜索、零样本分类、聚类、下游模型训练和研究等任务。数据集基于ReLAION-2B-en-research-safe，许可为CC-BY 4.0。

The ReLAION-2B Natural Embeddings dataset contains pre-computed CLIP ViT-H/14 embeddings for approximately 500 million natural photographs from ReLAION-2B. Each photograph has a naturalness score (0.7-1.0) indicating its degree of naturalness. The dataset structure includes image URLs, naturalness scores, row indices in the original LAION-2B embeddings, and 768-dimensional CLIP embedding vectors. The dataset is primarily used for tasks such as image similarity search, zero-shot classification, clustering, downstream model training, and research. The dataset is based on ReLAION-2B-en-research-safe and is licensed under CC-BY 4.0.