# dcvlm_pool_large by ML Foundations: AI Training Dataset Source: https://savrn.com/datasets/dcvlm-pool-large Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Dataset Card The raw candidate pool at the large scale of our DataComp-VLM benchmark: 1,949,321,868 samples / 166.7 TB across 166 source datasets, as WebDataset tar shards — ≈4× the medium pool. This pool is unfiltered and unmixed. It is the input to a data-curation experiment, not a training set. You choose the filters and the mixing ratios, and create another training set. If you instead want a ready-to-train dataset, use dcvlm-baseline-200b (our reference SoTA DCVLM-baseline dataset) or dcvlm-balanced-200b. One directory per source dataset, continuing the medium pool's shard numbering: Four directories are nested at this scale (the medium pool only needed this for datacomp1b). Their shard counts… Excerpt from the card by ML Foundations, licensed other. ## Details Repository mlfoundations/dcvlm_pool_large Publisher ML Foundations Task category Image and text to text Tags webdataset, multimodal, vision-language Size category 1B