FineVision Vision-Language Training Corpus

## Overview FineVision is a large-scale open vision-language dataset assembled by HuggingFaceM4 (Hugging Face's multimodal team) for training state-of-the-art open Vision-Language Models. It contains 17.3M images, 24.3M samples, 88.9M conversational turns, and approximately 9.5B answer tokens spread across many sub-configurations. Data is distributed as Parquet shards with image + text modalities, and is organized into many named subsets (each a different upstream source or task family) loadable independently via the `datasets` library. ## Schema Schema varies by subset, but typical columns include: - `images` — list[image] — one or more images associated with the sample - `texts` — list[dict] — multi-turn conversation, each turn with `user` / `assistant` (and sometimes `system`) fields - `source` — string — upstream dataset identifier - `id` / `sample_id` — string — stable sample identifier - additional task-specific fields per subset (e.g. bounding boxes, OCR text, captions) - +N more columns depending on subset ## Sources - HuggingFaceM4/FineVision on Hugging Face — https://huggingface.co/datasets/HuggingFaceM4/FineVision — see dataset card for license - Companion blog/space: https://huggingface.co/spaces/HuggingFaceM4/FineVision - Paper: arXiv:2510.17269 ## Methodology FineVision is a curated aggregation of many existing open vision-language datasets (captioning, VQA, OCR, document understanding, grounding, chart/diagram reasoning, multi-image reasoning, etc.) reformatted into a unified multi-turn conversational schema suitable for instruction tuning of VLMs. The HuggingFaceM4 team performed deduplication, quality filtering, and schema normalization across upstream sources. Each subset retains its provenance via the `source` field so users can trace examples back to upstream datasets and their original licenses. ## Known gaps & limitations Because FineVision aggregates many upstream datasets, per-subset licenses vary and downstream users should verify license compatibility for each subset they use. Coverage is heavily English-centric. Image domains skew toward existing academic VLM benchmarks (natural images, documents, charts) and may underrepresent specialized verticals (medical, satellite, industrial). The corpus has not been deduplicated against common multimodal eval suites (MMMU, MMBench, ChartQA, DocVQA, etc.), so benchmark leakage is a real risk when training models intended to be evaluated on those suites. Source does not document all gaps exhaustively; buyers should validate empirically. ## Intended use & out-of-scope - IS for: pretraining and instruction-tuning open Vision-Language Models, RAG over multimodal content, building multimodal agents, large-scale VLM research. - NOT for: benchmark training without leakage checks against MMMU/MMBench/ChartQA/DocVQA-style evals; production use in regulated domains (medical, legal) without domain-specific validation; tasks requiring guaranteed per-image licensing without per-subset license review. _Federated dataset: 9,155 parquet shards, 4177.30 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: FineVision — 24M-Sample Vision-Language Training Corpus Massive open VLM training set: 17.3M images, 24.3M samples, 88.9M turns, 9.5B answer tokens. Parquet, multi-subset, ready for fine-tuning state-of-the-art vision-language models.