Dataset · allenai
Synthetic questions on charts, tables, and documents
allenai/pixmo-docs
A synthetic dataset of question-answer pairs based on generated images of documents, tables, and charts.
What's inside
The dataset contains 100K to 1M question-answer pairs over computer-generated images including charts, tables, diagrams, and documents. Images are rendered from code written by the Claude model, while questions and answers are generated using GPT-4o mini. The AllenAI authors note they now recommend using the newer CoSyn-400k and CoSyn-point versions.
What you can do with it
- Train multimodal models (VLMs) to understand charts and diagrams
- Fine-tune visual question answering (Visual QA) systems on documents
- Tasks
- Visual question answering
- Size
- 100K–1M rows
- Downloads this month
- 6.6K
- Likes
- 35
Models trained on it
- DAMO-NLP-SG/VideoLLaMA3-7BVideo text → text 5.5K
- DAMO-NLP-SG/VideoLLaMA3-2BVideo text → text 2.6K
- DAMO-NLP-SG/VideoLLaMA3-2B-ImageVisual question answering 72
- DAMO-NLP-SG/VideoLLaMA3-7B-ImageVisual question answering 66
- henilchopada/laya-vVisual question answering 44
- cbipok/VideoLLaMA3-2B-forkVideo text → text 28
How to load it
from datasets import load_dataset
ds = load_dataset("allenai/pixmo-docs")