LLaVA-OneVision-2-Data
Video clips and captions for LLaVA-OneVision-2 mid-training.
Open dataset Updated 2026-06-18
What is LLaVA-OneVision-2-Data?
LLaVA-OneVision-2-Data provides video clips stored in WebDataset tar format together with separate JSONL caption files for 30 s and 60 s segments.
It supports model developers building or fine-tuning systems for video-text-to-text, visual question answering, and image-text-to-text workloads.
What you can build with LLaVA-OneVision-2-Data
Mid-train video-text models
Use the 60-second clips and paired captions to continue pretraining multimodal models on video-language alignment.
Build short-video reasoning pipelines
Train or evaluate models on the provided 30-second and 60-second caption data for temporal reasoning tasks.
Prepare WebDataset shards for large-scale training
Load the 10,809 shards directly into training loops that expect tar-based video and JSONL metadata.
Load LLaVA-OneVision-2-Data
Python
from datasets import load_dataset
ds = load_dataset("mvp-lab/LLaVA-OneVision-2-Data")- 1pip install datasets webdataset
- 2from datasets import load_dataset
- 3ds = load_dataset('mvp-lab/LLaVA-OneVision-2-Data')
- 4Iterate over the WebDataset shards to stream video clips
- 5Read the accompanying JSONL files for 30s and 60s captions
LLaVA-OneVision-2-Data: pros & cons
Pros
- +10,809 ready-to-stream WebDataset shards
- +Targeted video-text and reasoning content
- +Includes both 30-second and 60-second captions
- +Directly supports mid-training of LLaVA-style models
Cons
- –Intended only for mid-training, not full pretraining
- –Full download requires substantial storage
- –Tied to a specific model family and format
Did you find this helpful?
Frequently asked questions
A collection of video clips and captions released to train the LLaVA-OneVision-2 multimodal model family.
User reviews
Verified reviews from the community shape this listing's rating.
Loading reviews…
Similar datasets
Other images & vision options worth comparing.