NuTonic/oceanscout-sft-v1
收藏资源简介:
OceanScout SFT是一个基于Sentinel-2卫星图像的海事SFT样本数据集,主要用于目标检测和视觉语言模型微调(VLM-SFT)。数据集通过NDWI(归一化差异水指数)定义水域,并在水域上检测船只候选目标。每个训练样本包含单个场景的RGB图像块,并附带海事描述文本和检测到的目标区域(如果存在)。数据集分为训练集(1355条)、验证集(173条)和测试集(202条),总计1730条记录。数据集布局包括JSONL格式的训练文件、PNG图像和元数据JSON文件。数据集的生成和上传过程详细说明了所需的命令行参数和环境变量。
OceanScout SFT is a maritime SFT sample dataset based on Sentinel-2 satellite imagery, primarily used for object detection and vision-language model fine-tuning (VLM-SFT). The dataset defines water bodies using NDWI (Normalized Difference Water Index) and detects vessel candidates on water. Each training sample includes a single post-scene RGB chip and is accompanied by maritime caption text and detected regions (if any). The dataset is divided into train (1355 records), validation (173 records), and test (202 records) splits, totaling 1730 records. The dataset layout includes JSONL files for training, PNG images, and metadata JSON files. The README provides detailed instructions for dataset generation, upload, and required environment variables.




