allenai/olmoearth-paper-embeddings

该数据集包含来自26个地球观测基础模型的预提取嵌入，这些嵌入基于OlmoEarth论文中的Table 2，在24个下游任务上进行了评估。数据集存储为.pt文件，总大小约1.1 TB，支持分类和分割任务，包括训练、验证和测试分割。每个文件包含嵌入和标签张量，嵌入维度因模型而异（从128到4096），标签格式根据任务类型（单标签、多标签或分割）调整。数据集旨在方便下游实验（如KNN、线性探测或微调），无需重新提取嵌入，并附带论文最佳超参数设置以支持复现。数据集涵盖多种地球观测模态（如Sentinel-1、Sentinel-2、Landsat），并包括已知注意事项和存储细节。This dataset contains pre-extracted embeddings from 26 Earth-observation foundation models evaluated on the 24 downstream tasks that make up Table 2 of the OlmoEarth paper. The embeddings are saved as .pt files with a total size of approximately 1.1 TB, covering train, validation, and test splits for classification and segmentation tasks. Each file includes embeddings and labels, with embedding dimensions varying by model (from 128 to 4096) and label formats adapted to task types (single-label, multilabel, or segmentation). The dataset facilitates downstream experiments like KNN, linear probing, or fine-tuning without re-extraction, and includes paper-best hyperparameters for reproducibility. It spans multiple Earth-observation modalities (e.g., Sentinel-1, Sentinel-2, Landsat) and includes known caveats and storage details.