license: cc-by-4.0
language:
- en
task_categories:
- question-answering
tags:
- astronomy
- astrophysics
- ai4science
- llm-evaluation
- machine-learning
- method-selection
- benchmark
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path: benchmark.json
Astro2Method
A benchmark for evaluating LLMs on machine learning method selection in astronomy.
Astro2Method contains 500 expert-reviewed binary-choice questions drawn from 346 astronomy papers published between 2024/01/01 and 2026/06/30. Each question presents an astronomical problem and two candidate ML methods. The task is to select which method performs better for the given problem, based on the comparison reported in the source paper.
Dataset Overview
| Property | Value |
|---|---|
| Questions | 500 |
| Source papers | 346 |
| Easy | 376 |
| Hard | 124 |
| Publication years | 2024–2026 |
| Language | English |
| Format | JSON |
| License | CC BY 4.0 |
The dataset is provided as an evaluation set.
Benchmark Construction
We collected astronomy papers involving machine learning and identified method comparisons from reported experimental results. We retained comparisons with meaningful performance differences and explanations provided by the original authors.
Each comparison was converted into a binary-choice question and reviewed by human experts. Difficulty was assigned based on the completeness of the authors' explanations:
- Easy: The source paper directly and completely explains the performance difference.
- Hard: The source paper provides a partial explanation.
Comparisons without supporting explanations were excluded. A single source paper may contribute multiple questions.
Data Format
The dataset is stored in benchmark.json. Each record contains seven fields:
| Field | Description |
|---|---|
question_id |
Unique question identifier |
paper_id |
NASA ADS bibliographic code of the source paper |
question |
Astronomical problem and relevant conditions |
option_a |
First candidate method |
option_b |
Second candidate method |
correct_option |
Correct answer (A or B) |
difficulty |
easy or hard |
Usage
Load the dataset using Hugging Face Datasets:
from datasets import load_dataset
dataset = load_dataset("Xingzhongfan/Astro2Method", split="test")
print(dataset[0])
For evaluation, provide the question and two candidate methods to the model. Use correct_option as the ground truth.
Report accuracy separately for Overall, Easy, and Hard questions.
To avoid answer-position bias, use randomized option ordering and update the answer labels consistently. Keep the correct answers separate from model inputs.
License
Astro2Method is released under the Creative Commons Attribution 4.0 International License.
The license covers the original benchmark content. Source papers retain their respective copyrights.
Citation
If you use Astro2Method in your research, please cite:
Astro2Method: Benchmarking LLMs on Machine Learning Method Selection in Astronomy.
The BibTeX citation will be added following the arXiv release.