Dataset · OpenGVLab
Multimodal Corpus for VLM Training (InternVL3.5)
OpenGVLab/MMPR-v1.2
A large updated English image-text pair dataset used to improve InternVL3.5 models.
What's inside
The dataset contains between 1M and 10M English records for image-text-to-text tasks, with images packaged in a multi-volume images.zip archive. It is an expanded version of OpenGVLab's MMPR and MMPR-v1.1 corpora with added data sources to boost data diversity; generation prompts are provided separately.
What you can do with it
- SFT or preference training (MPO) of multimodal language models (VLMs)
- Improving reasoning accuracy and visual understanding in InternVL-class models
- Tasks
- Image text → text
- Languages
- English
- Size
- 1M–10M rows
- Downloads this month
- 2.2K
- Likes
- 42
Models trained on it
- OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview-HFImage text → text 666.2K
- OpenGVLab/InternVL3_5-8BImage text → text 106.7K
- OpenGVLab/InternVL3-1B-hfImage text → text 72.7K
- OpenGVLab/InternVL3-1BImage text → text 54.2K
- OpenGVLab/InternVL3-2B-hfImage text → text 51.3K
- mlx-community/gemma-3-4b-it-qat-4bitImage text → text 46.6K
- OpenGVLab/InternVL3-8BImage text → text 37.5K
- OpenGVLab/InternVL3-2BImage text → text 34.8K
How to load it
from datasets import load_dataset
ds = load_dataset("OpenGVLab/MMPR-v1.2")