lance-format/mnist-lance
收藏资源简介:
该数据集是经典MNIST手写数字数据集的Lance格式版本,包含70,000个28×28灰度数字图像,覆盖10个平衡类别(数字0-9)。每行数据包含内联PNG字节(图像数据)、数字标签(0-9的整数)、人类可读的类别名称(如“0”到“9”的字符串)以及预计算的CLIP图像嵌入(基于OpenCLIP ViT-B-32模型,512维,余弦归一化)。数据集采用Lance列式存储格式,支持高效随机访问和向量相似性搜索。关键特性包括:内联PNG字节无需外部文件、预计算CLIP嵌入、在标签列上构建的标量索引(BTREE和BITMAP)以及用于向量搜索的IVF_PQ索引。数据集分为训练集(60,000行)和测试集(10,000行),可通过HuggingFace datasets、LanceDB或Lance库加载,并支持搜索、筛选、数据演化、训练、版本控制和子集物化等操作。
This dataset is a Lance-formatted version of the classic MNIST handwritten digit dataset, containing 70,000 28×28 grayscale digital images spanning 10 balanced classes (digits 0 through 9). Each row includes inline PNG bytes (storing the raw image data), numeric labels (integers ranging from 0 to 9), human-readable class names (strings of "0" to "9"), and pre-computed CLIP image embeddings based on the OpenCLIP ViT-B-32 model: these embeddings are 512-dimensional and cosine-normalized. Built on the Lance columnar storage format, the dataset supports efficient random access and vector similarity search. Its key features are as follows: inline PNG bytes that require no external files, pre-computed CLIP embeddings, scalar indexes (BTREE and BITMAP) constructed on the label column, and an IVF_PQ index for vector search. The dataset is split into a training set (60,000 rows) and a test set (10,000 rows). It can be loaded via the HuggingFace Datasets, LanceDB, or Lance libraries, and supports various operations including search, filtering, data evolution, training, version control, and subset materialization.




