AstroMLab/USAAAO_QA

USA Astronomy and Astrophysics Organization (USAAAO) Open-ended Questionaire

This dataset contains structured question–answer pairs from the USAAAO (2017-2019), converted from pdf to LaTeX sources, into a machine-readable JSONL format.

The dataset includes:

• Short, medium, and long-form astrophysics problems
• Multi-part questions (grouped via parent_id)
• Full worked solutions
• Associated figures (JPEG images) where applicable

The goal of this dataset is to support evaluation and benchmarking of large language models (LLMs) on:

• Advanced scientific reasoning
• Mathematical derivation
• Multi-step physics problem solving
• Multimodal reasoning (text + figures)

---

Dataset Structure

Each record contains:

• id – unique question or sub-question identifier
• parent_id – shared ID for multi-part questions (null for single-part)
• year – exam year
• part – sub-part label (e.g., "a", "b")
• part_order – ordering index for sub-parts
• question_type – typically "open-ended"
• question_length – "short", "medium", or "long"
• question – problem statement
• answer – full solution text
• image – relative path to image (if applicable)
• caption – figure caption (if applicable)

Directory Layout

USAAAO_QA/
├── README.md
├── data/
│   ├── 2017.jsonl
│   ├── 2018.jsonl
│   └── 2019.jsonl
└── images/
    ├── 2017/
    │   ├── 2017-1.jpg
    │   ├── 2017-2.jpg
    │   └── ...
    ├── 2018/
    │   └── ...
    └── 2019/
        ├── 2019-1.jpg
        ├── 2019-2.jpg
        ├── 2019-3.jpg
        └── 2019-4.jpg

---

Loading the Dataset

If using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("AstroMLab/USAAAO_QA")

If loading locally from JSONL:

from datasets import load_dataset

dataset = load_dataset("json", data_files="data/2019.jsonl")

Citation

If you use this dataset, please cite:

@misc{usaaao_dataset_2025, author = {AstroMLab}, title = {USAAAO QA dataset}, year = {2025}, }