FineVision

FineVision is an open vision-language training mixture assembled by the M4 team at Hugging Face, the group behind the Idefics and SmolVLM models. It unifies more than 200 publicly available sources into 185 subsets, reporting roughly 24.3 million samples, 17.3 million images, 88.9 million conversation turns and 9.5 billion answer tokens. The mixture spans the tasks that general vision-language models need: visual question answering, optical character recognition, charts and tables, document understanding, captioning, and graphical user interface or agent navigation. Its distinguishing feature is transparency. Every turn was rated by 32B-parameter vision-language models across four qualitative metrics, and the corpus was decontaminated against common evaluation benchmarks, so you can see how each subset was judged and build your own training mixtures rather than trusting an opaque blend. You access it through the Hugging Face Hub as Parquet, loadable with the datasets library or Dask, and the per-subset structure lets you pull only the parts you want (individual subsets range from about 1GB to low tens of GB). For RAG and AI work it is primarily a supervised fine-tuning corpus for multimodal models rather than a retrieval index, but it is a strong base for training the encoders and instruction-followers that read documents, screenshots and charts inside a multimodal RAG pipeline. It is best for teams building or fine-tuning open VLMs who want auditable data provenance. The licence position needs care. There is no single blanket licence: the prompts are offered under CC BY 4.0 only where the authors hold rights, and each of the 200-plus upstream datasets keeps its own terms, several of which are CC BY-NC or research-only, while some subsets derive from proprietary-model outputs. Commercial use therefore requires per-subset checking. It sits alongside the image-text corpora we catalogue such as LAION-5B, OBELICS and DataComp, but is curated and task-labelled rather than raw web pairs.