FineVision
Hugging Face (M4 team)
17.3 million images and 24.3 million samples merged from many sources to train open vision-language models.
A large, curated collection of vision-language datasets: 88.9 million conversation turns and 9.5 billion answer tokens. One row is an image-and-text sample organized into subsets. It is designed for training open vision-language models and cleans, deduplicates, and standardizes the source sets.
Access
- Cost
- FreeNo charge; sources keep their own licenses.
- License
- Provider's own termsNot given as one tag; mixed per-subset licenses
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadParquet
What's in it
- One row is
- One row per image-text sample
- Covers
- Not geographic
- Years
- 2025
- Updated
- Irregular
- Size
- 17.3M images, 24.3M samples, 88.9M turns, 9.5B answer tokens
Quality
Documented curation with dedupe and test-set checks, but license terms vary by source subset.
- Per-subset licenses must be checked
- Some content is machine-generated
- Images drawn from many sources with different rights
Connects to
- The Cauldron: Predecessor of 50 datasets from the same group
- LLaVA-OneVision-1.5 Mid-Training 85M: Competing large open multimodal training set
Questions it could answer
- Which subsets supply most of the answer tokens?
- How much overlap exists with common vision benchmarks?