Skip to content

Dataset · facebook

Spatial reasoning benchmark for 3D shapes

facebook/sixthview_benchmark

A Meta (Facebook) 3D logic benchmark to predict the hidden face of a block of cubes using images and text.

What's inside

Contains 604 English spatial reasoning tasks featuring text instructions, 11,634 images, and 3D structure descriptions. Each example includes model input data in jsonl format and a reference response with the actual geometry of the assembled cube object.

What you can do with it

  • Evaluating spatial reasoning and 3D understanding in multimodal LLMs
  • Testing models' ability to mentally rotate and complete 3D objects
Languages
English
Size
1K–10K rows
Downloads this month
596
Likes
4

How to load it

from datasets import load_dataset

ds = load_dataset("facebook/sixthview_benchmark")