YeQingWen/FlowGuard-Dataset

FlowGuard数据集旨在支持FlowGuard框架的训练和评估，专注于安全感知的图像生成和检测。数据集包含9种不同的扩散/生成架构生成的图像，按模型架构组织，存储为大小平衡的tar分片，并过滤以保留必要的监督信号。数据存储格式为Parquet分片，按split、model和label组织。每个样本包含图像及其元数据，如模型名称、split、安全标签、case ID、step类型和step索引等。数据集质量通过自动审计和人工监督相结合的方式进行保证。

The FlowGuard Dataset is designed to support the training and evaluation of the FlowGuard framework, with a focus on safety-aware image generation and detection. The dataset includes generations from 9 different diffusion / generative architectures, organized by model architecture, packed into size-balanced tar shards, and filtered to retain only essential supervision signals. The dataset is stored as Parquet shards organized by split, model, and label. Each row contains one image sample with metadata such as model name, split, safety label, case ID, step type, and step index. The dataset quality is ensured through a hybrid pipeline of automatic auditing and strict human supervision.