MINT-1T

The first 1T-token multimodal interleaved dataset — 3.4B images together with text drawn from HTML, PDFs, and arXiv — roughly 10× the scale of prior open interleaved corpora (NeurIPS 2024). UW-led (Anas Awadalla, mlfoundations) with Salesforce, Stanford, UT Austin, and Berkeley collaborators; a core resource for training open multimodal models.