Dataset · jingyaogong
Bilingual image-text pairs for pre-training and SFT
jingyaogong/minimind-v_dataset
A large Chinese-English corpus based on ALLaVA-4V for training and fine-tuning vision-language models.
What's inside
It contains images (nature, documents, charts, synthetics) with detailed descriptions and instructions in Chinese and English from the LAION and VFLAN collections. The data is split into two parquet files: ~1.27 million records for pre-training (~640K unique images) and ~2.9 million for SFT (~650K unique images). Although automatic tags indicate a size under one thousand, the card itself describes volumes in the millions.
What you can do with it
- Initial alignment of visual and text tokens (VLM pre-training)
- Instruction fine-tuning (SFT) of multilingual models for image-based QA
- Training models for document, chart, and diagram recognition
- Tasks
- Visual question answering
- Languages
- Chinese, English
- Size
- up to 1K rows
- Downloads this month
- 1.2K
- Likes
- 39
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("jingyaogong/minimind-v_dataset")