Dataset · OpenGVLab
Large multimodal preference dataset for VQA
OpenGVLab/MMPR-v1.1
About 3 million examples for training visual reasoning and aligning VLM models.
What's inside
Large multimodal preference dataset for image-text reasoning (VQA) in English. Contains about 3 million examples, with images packaged in a multi-volume images.zip archive. The authors fine-tuned the InternVL2-8B model on this data.
What you can do with it
- Aligning (RLHF/DPO) multimodal language models for reasoning tasks
- Training VLMs to answer complex questions about images
- Comparing visual-language model quality on large data volumes
- Tasks
- Visual question answering
- Languages
- English
- Size
- 1M–10M rows
- Downloads this month
- 630
- Likes
- 46
Models trained on it
- OpenGVLab/InternVL2_5-4B-MPOImage text → text 245.8K
- OpenGVLab/InternVL2_5-8B-MPOImage text → text 875
- mradermacher/InternVL2_5-78B-MPO-i1-GGUF 494
- mradermacher/InternVL2_5-4B-MPO-GGUF 342
- mradermacher/InternVL2_5-78B-MPO-GGUF 332
- mradermacher/InternVL2_5-38B-MPO-i1-GGUF 282
- mradermacher/InternVL2_5-4B-MPO-i1-GGUF 231
- mradermacher/InternVL2_5-38B-MPO-GGUF 215
How to load it
from datasets import load_dataset
ds = load_dataset("OpenGVLab/MMPR-v1.1")