jonathan-roberts1/GRAB-real

GRAB: A Challenging GRaph Analysis Benchmark for Large Multimodal Models

GRAB consists of 3 splits: GRAB, GRAB-real and GRAB-lite. This is the dataset for GRAB-real.

Dataset Summary

The 'real' split of the GRAB benchmark is comprised of 1114 real questions covering 4 real tasks: noise, screenshots, whiteboard, sketches.

Large multimodal models (LMMs) have exhibited proficiencies across many visual tasks. Although numerous benchmarks exist to evaluate model performance, they increasingly have insufficient headroom and are unfit to evaluate the next generation of frontier LMMs.

To overcome this, we present GRAB, a challenging benchmark focused on the tasks human analysts might typically perform when interpreting figures. Such tasks include estimating the mean, intercepts or correlations of functions and data series and performing transforms.

We evaluate a suite of 20 LMMs on GRAB, finding it to be a challenging benchmark, with the current best model scoring just 21.0%.

Example usage

from datasets import load_dataset

# load dataset
grab_real_dataset = load_dataset("jonathan-roberts1/GRAB-real", split='GRAB_real')
"""
Dataset({
    features: ['pid', 'question', 'answer', 'decoded_image', 'real_task', 'grab_task', 'category', 'complexity', 'image'],
    num_rows: 1114
})
"""
# query individual questions
grab_real_dataset[40] # e.g., the 41st element
"""
{'pid': 40, 'question': 'Estimate the mean maximum value of the data series shown in the plot.
Give your answer to 1 decimal place.', 'answer': '6.9', 'decoded_image':
<PIL.PngImagePlugin.PngImageFile image mode=RGB size=2298x2209 at 0x1320586D0>,
'real_task': 'noise', 'grab_task': 'series', 'category': 'maximum value',
'complexity': 5, 'image': 'images/noise_images/40.jpg'}
"""
question_40 = grab_real_dataset[40]['question'] # question
answer_40 = grab_real_dataset[40]['answer'] # ground truth answer
pil_image_40 = grab_real_dataset[0]['decoded_image']

Note -- the 'image' feature corresponds to filepaths in the images dir in this repository: (https://huggingface.co/datasets/jonathan-roberts1/GRAB-real/tree/main/images.zip)

Please visit our GitHub repository for example inference code.

Dataset Curators

This dataset was curated by Jonathan Roberts, Kai Han, and Samuel Albanie

Citation Information

@inproceedings{roberts2025grab,
  title={GRAB: A challenging graph analysis benchmark for large multimodal models},
  author={Roberts, Jonathan and Han, Kai and Albanie, Samuel},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={1644--1654},
  year={2025}
}