LLaVA-OneVision-1.5-Mid-Training-85M
85M samples from eight vision datasets for LLaVA-OneVision-1.5 mid-training.
Open dataset Updated 2026-06-18
What is LLaVA-OneVision-1.5-Mid-Training-85M?
LLaVA-OneVision-1.5-Mid-Training-85M consists of image and text data compiled from the eight listed public sources for use in multimodal model training.
It is intended for researchers developing or reproducing open multimodal large language models that require large-scale mid-training stages.
What you can build with LLaVA-OneVision-1.5-Mid-Training-85M
Pre-train vision-language models
Use the 85M aggregated samples to continue pre-training models like LLaVA-OneVision-1.5 on diverse image-text pairs from multiple public sources.
Benchmark data scaling experiments
Measure the impact of mid-training on model performance by subsampling this 10-100M dataset and comparing against smaller curated sets.
Build multimodal retrieval systems
Leverage the combined ImageNet-21k, SA-1B, and web-scale sources to train or evaluate image-text retrieval components.
Load LLaVA-OneVision-1.5-Mid-Training-85M
Python
from datasets import load_dataset
ds = load_dataset("mvp-lab/LLaVA-OneVision-1.5-Mid-Training-85M")- 1pip install datasets
- 2from datasets import load_dataset
- 3ds = load_dataset('mvp-lab/LLaVA-OneVision-1.5-Mid-Training-85M')
- 4Access the 'train' split and iterate over image-text pairs
- 5Filter or subsample as needed for your training pipeline
LLaVA-OneVision-1.5-Mid-Training-85M: pros & cons
Pros
- +Large scale (85M samples) from eight established sources
- +Released specifically for LLaVA-OneVision-1.5 mid-training
- +Directly loadable via Hugging Face datasets library
- +Covers broad visual domains including segmentation and captioning data
Cons
- –Size requires substantial storage and compute
- –Quality and license consistency varies across source datasets
- –No built-in train/validation split provided
Did you find this helpful?
Frequently asked questions
An 85-million-sample vision dataset aggregated from ImageNet-21k, LAIONCN, DataComp-1B and other public collections to support mid-training of the LLaVA-OneVision-1.5 framework.
User reviews
Verified reviews from the community shape this listing's rating.
Loading reviews…
Similar datasets
Other images & vision options worth comparing.