Dataset · mlfoundations
121M Example Raw Pool for VLM Data Curation
mlfoundations/dcvlm_pool_small
Unfiltered 10.3 TB corpus from 166 sources for VLM data filtering experiments.
What's inside
This is a small candidate pool from the DataComp-VLM benchmark, comprising 120.9 million image-text pairs (approximately 187.5 billion tokens and 10.3 TB) across six languages, including English, Chinese, and German. Data is collected from 166 sources and packed into WebDataset tar shards. The pool is unfiltered and unbalanced, intended for developing data selection algorithms rather than direct training.
What you can do with it
- Developing and verifying multimodal data filtering and deduplication algorithms
- Determining optimal source mixing ratios for VLM training
- Reproducing DataComp-VLM benchmark experiments
- Tasks
- Image text → text, Visual question answering
- Languages
- English, Chinese, German, French, Japanese, Korean, Latin, Finnish and 4 more
- Size
- 100M–1B rows
- Downloads this month
- 32.6K
- Likes
- 0
How to load it
from datasets import load_dataset
ds = load_dataset("mlfoundations/dcvlm_pool_small")