Skip to content

Dataset · google

Google DeepMind benchmark for multimodal agents

google/multimodal-knowledge-agent

10 demonstration tasks from Google DeepMind for evaluating AI agents working with complex graphics and documents.

What's inside

A small set (less than 1,000 examples) in English, including images, documents, and geospatial data. In each task, the agent is given a workspace of multiple files with key visual artifacts like architectural CAD drawings that cannot be solved with simple OCR.

What you can do with it

  • Evaluate multimodal AI agents on multi-step professional tasks
  • Test models in visual analysis scenarios of drawings and documents (VQA, Document QA)
Tasks
Image text → text, Visual question answering, Document question answering
Languages
English
Size
up to 1K rows
Downloads this month
8.8K
Likes
103

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("google/multimodal-knowledge-agent")