RealSyn100m
RealSyn100m is a very large-scale multimodal dataset combining real images and associated, synthetic or extracted texts. It is designed to train effective and diverse vision-language models.
Up to 100 million examples, including images and associated texts, several formats used depending on the splits
MIT
Description
RealSyn100m is a vision-language dataset containing up to 100 million examples of images associated with realistic and synthetic texts. It uses an advanced pipeline for extracting real world data and synthetic text generation to improve semantic diversity and richness.
What is this dataset for?
- Pre-train or refine large-scale multimodal models (e.g. CLIP and variants)
- Improve the detailed understanding and representation of images with varied texts
- Allow better generalization on rare concepts thanks to semantic balancing
Can it be enriched or improved?
The dataset can be enriched by adding new sources of images or texts, or by generating synthetic texts adapted to specific contexts. The additional annotation can refine the granularity of image-text associations.
🔎 In summary
🧠 Recommended for
- AI research laboratories
- Businesses with big computing
- Advanced vision-language projects
🔧 Compatible tools
- Hugging Face Datasets
- PyTorch
- TensorFlow
- High-volume management tools (Spark, Dask)
💡 Tip
Provide a distributed storage and processing system to effectively manage the dataset.
Frequently Asked Questions
Is this dataset suitable for training on modest resources?
No, its size requires efficient infrastructures and a lot of computing resources.
What is the particularity of synthetic texts in RealSyn100m?
They are generated to increase semantic diversity and improve the coverage of rare concepts.
Can this dataset be used for non-vision-language tasks?
This dataset is specifically designed for vision-language tasks, but its images and texts can be extracted for other uses with adaptation.
