Datasets and downloadable files.
Questions written by crowd workers about Wikipedia paragraphs, where each answer is a span of the passage. Version 2.0 adds more than 50,000 questions that cannot be answered from the text, so models must also learn…
Volunteers record themselves reading short sentences, and other volunteers validate the clips. Each clip has a transcript and optional speaker details such as age range and accent. It is a key resource for speech…
English web pages from Common Crawl, filtered and deduplicated to keep higher-quality text. Smaller random samples (10B, 100B and 350B tokens) are provided for experiments that do not need the full corpus…
More than 100 million tokens from Good and Featured Wikipedia articles, kept as whole articles so long-range context is preserved. Comes in a tokenized version and a raw version. MaintainerSalesforce ResearchLicenseCC…
About 9 million images labelled with image-level classes, plus millions of bounding boxes over 600 classes, segmentation masks, visual relationships and localized narratives. MaintainerGoogleLicenseAnnotations CC BY…
About 15,000 instruction and response pairs written by Databricks employees. It covers brainstorming, classification, open and closed QA, summarization, information extraction and creative writing, and is used to…
150 flowers from three iris species, each measured on sepal and petal length and width. Small and clean, it is a standard first example for classification, clustering and plotting. MaintainerUCI Machine Learning…
70,000 grayscale 28x28 images of clothing in 10 classes (60k train, 10k test). It uses the same file format as MNIST, so code written for MNIST works unchanged. MaintainerZalando ResearchLicenseMITFormatGzipped IDX…