README.md

metadata

task_categories:
  - question-answering
tags:
  - science
pretty_name: Scientific Figure Interpretation Benchmark
size_categories:
  - 1k<n<10k
language:
  - en
configs:
  - config_name: default
    data_files:
      - split: CS_Figure2Caption
        path: data/CS_Figure2Caption-*
      - split: CS_Caption2Figure
        path: data/CS_Caption2Figure-*
      - split: General_Figure2Caption
        path: data/General_Figure2Caption-*
      - split: General_Caption2Figure
        path: data/General_Caption2Figure-*
dataset_info:
  features:
    - name: ID
      dtype: int64
    - name: Question
      dtype: string
    - name: Options
      sequence: string
    - name: Answer
      dtype: string
    - name: Category
      dtype: string
    - name: Images
      sequence: image
  splits:
    - name: CS_Figure2Caption
      num_bytes: 22992276
      num_examples: 500
    - name: CS_Caption2Figure
      num_bytes: 122043099
      num_examples: 500
    - name: General_Figure2Caption
      num_bytes: 290333873
      num_examples: 500
    - name: General_Caption2Figure
      num_bytes: 1475930020
      num_examples: 500
  download_size: 926209658
  dataset_size: 1911299268

SciFIBench

Jonathan Roberts, Kai Han, Neil Houlsby, and Samuel Albanie

NeurIPS 2024

Note: This repo has been updated to add two splits ('General_Figure2Caption' and 'General_Caption2Figure') with an additional 1000 questions. The original version splits are preserved and have been renamed as follows: 'Figure2Caption' -> 'CS_Figure2Caption' and 'Caption2Figure' -> 'CS_Caption2Figure'.

Dataset Description

• Homepage: SciFIBench
• Paper: SciFIBench: Benchmarking Large Multimodal Models for Scientific Figure Interpretation
• Repository SciFIBench

Dataset Summary

The SciFIBench (Scientific Figure Interpretation Benchmark) contains 2000 multiple-choice scientific figure interpretation questions covering two tasks. Task 1: Figure -> Caption involves selecting the most appropriate caption given a figure; Task 2: Caption -> Figure involves the opposite -- selecting the most appropriate figure given a caption. This benchmark was curated from the SciCap and ArxivCap datasets, using adversarial filtering to obtain hard negatives. Human verification has been performed on each question to ensure high-quality, answerable questions.

Example Usage

from datasets import load_dataset

# load dataset
dataset = load_dataset("jonathan-roberts1/SciFIBench") # optional: set cache_dir="PATH/TO/MY/CACHE/DIR"
# there are 4 dataset splits, which can be indexed separately 
# cs_figure2caption_dataset = load_dataset("jonathan-roberts1/SciFIBench", split="CS_Figure2Caption")
# cs_caption2figure_dataset = load_dataset("jonathan-roberts1/SciFIBench", split="CS_Caption2Figure")
# general_figure2caption_dataset = load_dataset("jonathan-roberts1/SciFIBench", split="General_Figure2Caption")
# general_caption2figure_dataset = load_dataset("jonathan-roberts1/SciFIBench", split="General_Caption2Figure")
"""
DatasetDict({
    CS_Caption2Figure: Dataset({
        features: ['ID', 'Question', 'Options', 'Answer', 'Category', 'Images'],
        num_rows: 500
    })
    CS_Figure2Caption: Dataset({
        features: ['ID', 'Question', 'Options', 'Answer', 'Category', 'Images'],
        num_rows: 500
    })
    General_Caption2Figure: Dataset({
        features: ['ID', 'Question', 'Options', 'Answer', 'Category', 'Images'],
        num_rows: 500
    })
    General_Figure2Caption: Dataset({
        features: ['ID', 'Question', 'Options', 'Answer', 'Category', 'Images'],
        num_rows: 500
    })
})
"""

# select task and split
cs_figure2caption_dataset = dataset['CS_Figure2Caption']
"""
Dataset({
    features: ['ID', 'Question', 'Options', 'Answer', 'Category', 'Images'],
    num_rows: 500
})
"""

# query items
cs_figure2caption_dataset[40] # e.g., the 41st element
"""
{'ID': 40,
 'Question': 'Which caption best matches the image?',
 'Options': ['A)  ber vs snr for fft size=2048 using ls , lmmse , lr-lmmse .',
  'B)  ber vs snr for fft size=1024 using ls , lmmse , lr-lmmse algorithms .',
  'C)  ber vs snr for fft size=512 using ls , lmmse , lr-lmmse algorithms .',
  'D)  ber vs snr for fft size=256 using ls , lmmse , lr-lmmse algorithms with a 16 qam modulation .',
  'E)  ber vs snr for a bpsk modulation .'],
 'Answer': 'D',
 'Category': 'other cs',
 'Images': [<PIL.PngImagePlugin.PngImageFile image mode=RGB size=501x431>]}
"""

Source Data

More information regarding the source data can be found at: https://github.com/tingyaohsu/SciCap and https://mm-arxiv.github.io/.

Dataset Curators

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

Citation Information

@article{roberts2024scifibench,
  title={SciFIBench: Benchmarking Large Multimodal Models for Scientific Figure Interpretation},
  author={Roberts, Jonathan and Han, Kai and Houlsby, Neil and Albanie, Samuel},
  journal={arXiv preprint arXiv:2405.08807},
  year={2024}
}