Skip to content

Dataset · dwzhu

Benchmark for generating scientific illustrations from text

dwzhu/PaperBananaBench

ICML 2026 dataset for evaluating systems that turn research paper text into diagrams and charts.

What's inside

Created by the authors of PaperBanana (ICML 2026, spotlight) to test AI's ability to generate publication-ready methodology diagrams and statistical charts from paper text. Contains 1 to 10 thousand examples in English combining text and images. Includes two test tasks for translating text descriptions into illustrations.

What you can do with it

  • Evaluating multimodal and text-to-image models on scientific data
  • Developing AI assistants for formatting scientific publications
  • Comparing generators of complex diagrams and charts
Tasks
Text → image
Languages
English
Size
1K–10K rows
Downloads this month
3.4K
Likes
32

How to load it

from datasets import load_dataset

ds = load_dataset("dwzhu/PaperBananaBench")