metadata

license: cc0-1.0
tags:
  - science
  - physics
  - chemistry
  - problem-solving
  - multimodal
  - scientific-reasoning
  - bilingual
features:
  - name: problem_id
    dtype: string
  - name: discipline
    dtype: ClassLabel
    names:
      - Physics
      - Chemistry
  - name: problem
    features:
      - name: en
        dtype: string
      - name: zh
        dtype: string
  - name: answer
    dtype: string
  - name: image
    dtype: image

ScienceOlympiad: Challenging AI with Olympiad-Level Multimodal Science Problems

Dataset Description
Data Fields
Data Splits
Dataset Creation
How to Use
Citation
License

Dataset Description

The ScienceOlympiad dataset is a meticulously curated benchmark designed to test the limits of current AI models in scientific reasoning. It comprises elite, competition-level problems in physics and chemistry. Addressing the need for more diverse and realistic challenges, ScienceOlympiad introduces multimodal integration as a key dimension. Unlike purely text-based datasets, a significant portion of problems requires models to analyze and interpret visual information from diagrams and figures—a critical skill for real-world scientific problem-solving. Our objective is for ScienceOlympiad to serve as a pivotal evaluation tool, accelerating progress in domain-specific reasoning for the physical and chemical sciences while advancing the frontier of integrated text and visual understanding.

Data Fields

Each entry in the dataset represents a single problem and contains the following fields:

problem_id: (string) - A unique identifier for each problem.
discipline: (ClassLabel) - The scientific discipline of the problem (Physics or Chemistry).
problem: (A dictionary of string) - A dictionary containing the problem statement in both English (en) and Chinese (zh).
answer: (string) - The correct answer or solution.
image: (Image) - An image or diagram associated with the problem, stored in the images/ directory. This field will be None if no image is present.

Data Splits

This dataset consists of a single test split, provided in the test.parquet file.

Dataset Creation

The dataset is composed of a curated collection of problems from three sources: classic problems, adaptations of established problems, and original compositions. Each problem has been meticulously reviewed to ensure rigorous difficulty and unambiguous wording.

How to Use

You can load the dataset using the Hugging Face datasets library:

from datasets import load_dataset, Image
import matplotlib.pyplot as plt
import datasets

# Force a redownload to get the latest metadata
features = datasets.Features({
    'problem_id': datasets.Value('string'),
    'discipline': datasets.ClassLabel(names=['Physics', 'Chemistry']),
    'problem': {
        'en': datasets.Value('string'),
        'zh': datasets.Value('string'),
    },
    'answer': datasets.Value('string'),
    'image': datasets.Image(decode=True),
})
ds = load_dataset("ByteDance-Seed/ScienceOlympiad", features=features)
test_ds = ds['test']

example = test_ds[60]
image_data = example['image']


if image_data:
    plt.imshow(image_data)
    plt.show()

Citation

If you use the ScienceOlympiad dataset in your research or work, please consider citing it:

@misc{bytedance_seed_2025_scienceolympiad,
  author       = {[Bytedance-Seed]},
  title        = {ScienceOlympiad: Challenging AI with Olympiad-Level Multimodal Science Problems},
  year         = {2025},
  publisher    = {Hugging Face},
  journal      = {Hugging Face repository},
  howpublished = {\url{[https://huggingface.co/datasets/Bytedance-Seed/ScienceOlympiad](https://huggingface.co/datasets/Bytedance-Seed/ScienceOlympiad)}},
}

License

ScienceOlympiad is released under the CC0 1.0 Universal (CC0 1.0) Public Domain Dedication.

This means the work has been dedicated to the public domain by waiving all rights to the work worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law. You can copy, modify, distribute, and perform the work, even for commercial purposes, all without asking permission. For more details, see the LICENSE file or the full legal text of the CC0 license.

Datasets:

ByteDance-Seed
/

ScienceOlympiad