LVD-1689M is a large curated web-image dataset used by the DINOv3 authors for self-supervised pretraining. According to the DINOv3 paper and the model README, LVD-1689M contains approximately 1,689 million (1.689B) images sampled from a much larger pool (~17 billion) of web images collected from public Instagram posts. The authors describe LVD-1689M as a curated subset intended for large-scale SSL pretraining (used in the DINOv3 pretraining mixture); the paper and associated model documentation state the subset was created via clustering and balanced sampling to improve diversity and downstream generalization. LVD-1689M is not listed as a standalone public dataset on Hugging Face; primary references are the DINOv3 paper (arXiv:2508.10104), the Meta AI DINOv3 project page, the facebookresearch/dinov3 GitHub repo, and multiple DINOv3 model cards on Hugging Face that state models were pretrained on "LVD-1689M".
No results indexed yet — be the first to submit a score.
Submit a checkpoint and a reproduction script. We will run it, publish the score, and — if it takes the top — annotate the step on the progress chart with your name.