Voxel51/curated-mnist

curated-mnist数据集是对经典MNIST数据集的精选版本，包含了不同模型的预测结果、嵌入数据和多种分析字段。数据集分为训练集、验证集和测试集，其中训练集又进一步分为训练和验证两个子集。数据集包含了来自CLIP、LeNet和重新训练的LeNet模型的预测结果，并提供了图像嵌入和模型分析字段，如难度、误报率、独特性和代表性。数据集旨在用于教育目的，帮助用户探索模型比较、错误分析和以数据为中心的AI技术。

The curated-mnist dataset is a curated version of the classic MNIST dataset, enriched with model predictions, embeddings, and various analytical fields generated using the FiftyOne library. It contains the original 60,000 training and 10,000 test samples, with the training set further split into training and validation subsets. It includes predictions from OpenAIs CLIP model (zero-shot), a custom-trained LeNet-5 model, and a final LeNet-5 model that was fine-tuned on an augmented set of challenging samples. This dataset is ideal for educational purposes, allowing users to explore model comparison, error analysis, and data-centric AI techniques.