question_id string | paper_id string | question string | option_a string | option_b string | correct_option string | difficulty string |
|---|---|---|---|---|---|---|
Q0001 | 2024MNRAS.535.1844L | How can one obtain accurate photometric redshift estimates for high-redshift galaxies using only flux measurements from different observational filters as input? | MLP | LSTM | B | hard |
Q0002 | 2024ApJ...977...19D | How can the global 21-cm signal be reliably modeled from a specified set of cosmological and astrophysical parameters? | Two-layer LSTM | VAE | A | easy |
Q0003 | 2024ApJ...977...19D | How can the global 21-cm signal be reliably modeled from a specified set of cosmological and astrophysical parameters? | Two-layer LSTM | Single MLP | A | easy |
Q0004 | 2024ApJ...977...19D | How can the global 21-cm signal be reliably modeled from a specified set of cosmological and astrophysical parameters? | Two-layer LSTM | Multi-MLP + PCA + Decision Tree | A | easy |
Q0005 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | MLP | Neural density field | B | easy |
Q0006 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | vanilla PINN | Neural density field | B | easy |
Q0007 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | Transformer-PINN | Neural density field | B | easy |
Q0008 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | Neural density field | Self-adaptive PINN | A | hard |
Q0009 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | Self-adaptive PINN | MLP | A | easy |
Q0010 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | Self-adaptive PINN | vanilla PINN | A | easy |
Q0011 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | Self-adaptive PINN | Transformer-PINN | A | easy |
Q0012 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | MLP | Transformer-PINN | B | easy |
Q0013 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | Transformer-PINN | vanilla PINN | A | easy |
Q0014 | 2024AJ....168..242A | How can the near-surface gravity field of an irregular, nonuniform small body be accurately reconstructed from sampled simulated gravity-field data? | vanilla PINN | MLP | A | easy |
Q0015 | 2024A&A...692A.132H | How can we accurately and precisely measure time delays from simulated microlensed light curves of strongly lensed Type Ia supernovae? | Random Forest | LSTM + fully connected neural network | B | easy |
Q0016 | 2024A&A...692A..72C | For identifying galaxy-scale strong lenses in multiband survey images, which CNN architecture gives higher recall with very few false positives? | CNN | ResNet | B | easy |
Q0017 | 2024A&A...692A..72C | For identifying galaxy-scale strong lenses in multiband survey images, which CNN architecture gives higher recall with very few false positives? | Group-equivariant CNN | CNN | A | easy |
Q0018 | 2024A&A...692A..72C | How can we identify galaxy-scale strong gravitational lenses in multi-band survey images while maximizing recall at zero or very low contamination? | Group-equivariant CNN | ResNet | B | hard |
Q0019 | 2024ApJ...976..143T | For quasar spectra, how can the unabsorbed continuum in the Lyα forest be more accurately predicted from only spectral information redward of Lyα? | PCA | CNN | B | easy |
Q0020 | 2024ApJ...976..143T | For quasar spectra, how can the unabsorbed continuum in the Lyα forest be more accurately predicted from only spectral information redward of Lyα? | CNN | Non-Negative Matrix Factorization | A | easy |
Q0021 | 2024ApJ...976...37W | In cosmological hydrodynamical simulations, how can galaxy stellar masses be more accurately predicted from dark-matter subhalo properties and large-scale environment? | Explainable Boosting Machines | E(3)-invariant GNN | B | easy |
Q0022 | 2024A&A...691A.338C | For Solar System survey tracklets, how can interstellar objects be more reliably distinguished from other Solar System objects while minimizing false positives and false negatives? | Neural Network | Gradient Boosting Machines | B | easy |
Q0023 | 2024A&A...691A.338C | For Solar System survey tracklets, how can interstellar objects be more reliably distinguished from other Solar System objects while minimizing false positives and false negatives? | Random Forest | Neural Network | A | hard |
Q0024 | 2024A&A...691A.223V | Using low-resolution optical spectra, how can hot subdwarf binaries be more accurately identified against single hot subdwarfs? | CNN | Self-Organizing Map | A | easy |
Q0025 | 2024A&A...691A.223V | Using optical photometry and low-resolution spectrophotometric observations, which approach more accurately identifies hot subdwarf binaries against single hot subdwarfs? | weighted SVM | CNN | B | easy |
Q0026 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | Bayesian CNN | XGBoost | A | easy |
Q0027 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | Bayesian CNN | Bayesian MLP | A | hard |
Q0028 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | Bayesian CNN | MLP | A | easy |
Q0029 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | SVM | Bayesian CNN | B | easy |
Q0030 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | XGBoost | CNN | B | easy |
Q0031 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | MLP | CNN | B | easy |
Q0032 | 2024ApJ...974..159J | Using five-band galaxy images and photometry, how can photometric redshifts be estimated more reliably for weak-lensing cosmology? | SVM | CNN | B | easy |
Q0033 | 2024ApJ...974...96S | For astrometric candidate pairs, how can true astrometric binaries be distinguished from visual alignments with high discriminating power using proper motions, parallaxes, and angular and physical separations? | Logistic regression | Artificial neural network | B | easy |
Q0034 | 2024ApJ...973..163Y | For compact-binary phase-space trajectories learned from numerical simulations, how can orbital evolution be predicted more accurately and efficiently over longer prediction horizons? | Temporal Convolutional Network | LSTM | B | hard |
Q0035 | 2024ApJ...973..154H | How can stellar evolutionary tracks and observables be emulated more accurately from grids of stellar evolution models? | Conditional Normalizing Flows | Nearest-Neighbor Interpolation | A | easy |
Q0036 | 2024A&A...690A.224L | How can QSO classifications based on broadband colors achieve higher precision in a joint star-galaxy-QSO classification task? | hierarchical density-based clustering | sharpened-dimensionality-reduction + k-nearest-neighbor classification | B | easy |
Q0037 | 2024RAA....24i5023F | For stellar spectra represented as wavelet images, how can spectral types be classified more accurately? | Vision Transformer | feature-pyramid ConvNeXt-V2 CNN | B | easy |
Q0038 | 2024RAA....24i5023F | For stellar spectra represented as wavelet images, how can spectral types be classified more accurately? | Random Forest | feature-pyramid ConvNeXt-V2 CNN | B | easy |
Q0039 | 2024RAA....24i5023F | For stellar spectra represented as wavelet images, how can spectral types be classified more accurately? | stochastic-gradient-descent linear classifier | feature-pyramid ConvNeXt-V2 CNN | B | easy |
Q0040 | 2024RAA....24i5022M | For sunspot groups observed in white-light and magnetogram images, how can their magnetic classes be classified more reliably when edge dimming weakens image features? | dual-stream ResNet50 | dual-stream ResNet50 with channel-and-spatial attention | B | easy |
Q0041 | 2024RAA....24i5022M | For sunspot groups observed in white-light and magnetogram images, how can their magnetic classes be classified more reliably when edge dimming weakens image features? | dual-stream ResNet50 with squeeze-and-excitation attention | dual-stream ResNet50 with channel-and-spatial attention | B | easy |
Q0042 | 2024MNRAS.533..525M | Across combined real-and-simulated strong-lens test sets, which neural-network design gives more stable overall lens-classification performance? | five-layer classical CNN | ResNet-based CNN | A | hard |
Q0043 | 2024MNRAS.533..525M | For strong-lens searches evaluated on real lens candidates and real non-lenses, which neural-network design gives a better precision-recall trade-off at the adopted survey-search threshold? | ResNet-based CNN | block-structured CNN with batch normalization | A | hard |
Q0044 | 2024ApJS..274....7M | For low-to-high-quality galaxy-image super-resolution, which approach produces more visually realistic galaxy reconstructions? | SRGAN | EDSR | A | easy |
Q0045 | 2024ApJS..274....7M | For low-to-high-quality galaxy-image super-resolution, which approach produces more visually realistic galaxy reconstructions? | SRGAN | RCAN | A | easy |
Q0046 | 2024ApJS..274....7M | For low-to-high-quality galaxy-image super-resolution, which approach produces more visually realistic galaxy reconstructions? | SRGAN | efficient nonlocal contrastive attention network | A | easy |
Q0047 | 2024ApJS..274....7M | For low-to-high-quality galaxy-image super-resolution, which approach better matches high-resolution reference images in pixel-level and structural fidelity? | efficient nonlocal contrastive attention network | SRGAN | A | hard |
Q0048 | 2024ApJ...973...41Z | For Bayesian inference of reionization parameters from pure-signal 3D 21 cm light-cone images, which summary-compression approach yields tighter posterior constraints? | solid harmonic wavelet scattering transform + DELFI | 3D CNN + DELFI | A | easy |
Q0049 | 2024ApJ...973...41Z | For Bayesian inference of reionization parameters from 3D 21 cm light-cone images, which summary-compression approach yields tighter posterior constraints? | power spectrum + DELFI | solid harmonic wavelet scattering transform + DELFI | B | easy |
Q0050 | 2024ApJ...972....7R | For bright-transient survey candidates, which architecture better supports automated spectroscopic follow-up when source-level completeness and purity are prioritized? | multi-modal CNN | fully connected neural network | A | hard |
Q0051 | 2024ApJ...972....7R | For identifying bright extragalactic transients in survey alerts, which ML pipeline gives higher classification accuracy? | CNN-MLP fusion of images with photometric and catalogue features | CNN using images | A | easy |
Q0052 | 2024ApJ...972....7R | For classifying individual bright-transient survey alert packets before source-level follow-up policies are applied, which unimodal architecture gives higher alert-level accuracy and ROC-AUC? | fully connected neural network | uni-modal CNN | A | hard |
Q0053 | 2024A&A...689A.181M | For mock galaxy-cluster velocity-dispersion maps, which U-Net variant better recovers the radial ICL-fraction structure while reducing spurious small-scale ICL features? | Attention U-Net | U-Net | A | easy |
Q0054 | 2024A&A...689A.166C | For emission-line galaxy target selection across the tested photometric configurations, which classifier more consistently maintains higher target precision at high target recall? | Self-Organizing Map | MLP | B | easy |
Q0055 | 2024A&A...689A.166C | For emission-line galaxy target selection across the tested photometric configurations, which classifier more consistently maintains higher target precision at high target recall? | Self-Organizing Map | SVM | B | easy |
Q0056 | 2024A&A...689A.166C | For emission-line galaxy target selection across the tested photometric configurations, which classifier more consistently maintains higher target precision at high target recall? | Self-Organizing Map | Extra Trees Classifier | B | hard |
Q0057 | 2024A&A...689A.166C | For emission-line galaxy target selection across the tested photometric configurations, which classifier more consistently maintains higher target precision at high target recall? | Random Forest | Self-Organizing Map | A | hard |
Q0058 | 2024A&A...689A.166C | For emission-line galaxy target selection across the tested photometric configurations, which classifier more consistently maintains higher target precision at high target recall? | AdaBoost | MLP | B | hard |
Q0059 | 2024A&A...689A.166C | For emission-line galaxy target selection across the tested photometric configurations, which classifier more consistently maintains higher target precision at high target recall? | SVM | AdaBoost | A | hard |
Q0060 | 2024A&A...689A.143G | For faint companion spectra cross-correlated with multiple molecular templates, which classifier detects planets more sensitively at low false discovery rate when atmospheric characteristics are uncertain? | 1D CNN | Perceptron | A | hard |
Q0061 | 2024MNRAS.532.3322A | For identifying multi-component radio sources, which ML approach gives higher classification accuracy? | CNN using radio images | CNN-ANN fusion of radio images and catalogue features | B | easy |
Q0062 | 2024MNRAS.532.3322A | For identifying multi-component radio sources, which ML approach gives higher classification accuracy? | ANN using catalogue features | CNN-ANN fusion of radio images and catalogue features | B | easy |
Q0063 | 2024ApJS..273...27S | For reconstructing the cosmic distance ladder from Type Ia supernova apparent magnitudes with full covariance information, which method better predicts distances while preserving a smooth and monotonic redshift-distance relation? | Support Vector Regression | covariance-aware LSTM | B | hard |
Q0064 | 2024ApJS..273...27S | For reconstructing the cosmic distance ladder from Type Ia supernova apparent magnitudes with full covariance information, which method better predicts distances while preserving a smooth and monotonic redshift-distance relation? | MLP | covariance-aware LSTM | B | easy |
Q0065 | 2024ApJS..273...27S | For reconstructing the cosmic distance ladder from Type Ia supernova apparent magnitudes with full covariance information, which method better predicts distances while preserving a smooth and monotonic redshift-distance relation? | kNN regression | covariance-aware LSTM | B | easy |
Q0066 | 2024ApJ...970..120D | For extracting sunspot components from heterogeneous historical sunspot drawings when labels are available from one station and other stations are unlabeled, which segmentation approach gives more accurate cross-station segmentation? | Hybrid Task Cascade R-CNN | semisupervised dynamic mutual training with boundary-guided segmentation | B | easy |
Q0067 | 2024ApJ...970..120D | For extracting sunspot components from heterogeneous historical sunspot drawings when labels are available from one station and other stations are unlabeled, which segmentation approach gives more accurate cross-station segmentation? | semisupervised dynamic mutual training with boundary-guided segmentation | scale-aware attention Mask R-CNN with PointRend | A | easy |
Q0068 | 2024ApJ...970..120D | For extracting sunspot components from heterogeneous historical sunspot drawings when labels are available from one station and other stations are unlabeled, which segmentation approach gives more accurate cross-station segmentation? | semisupervised dynamic mutual training with boundary-guided segmentation | dynamic mutual training with DeepLab-v3 | A | easy |
Q0069 | 2024A&A...688A.199P | For compressing cosmic 21 cm lightcones for astrophysical parameter inference across the prior volume, which summary approach yields the tightest typical Fisher-forecast constraints? | 2D power spectrum + information-maximizing neural network | RNN | A | easy |
Q0070 | 2024A&A...688A.199P | Among neural-network summaries of cosmic 21 cm lightcones, which approach gives tighter typical astrophysical parameter constraints across the prior volume? | Information-maximizing neural network | RNN | A | hard |
Q0071 | 2024MNRAS.531.4363S | For separating accreted from in-situ stars in Milky Way-like galaxies when a classifier trained on one cosmological simulation suite must generalize to another, which method is less sensitive to domain shift? | MLP | XGBoost | A | hard |
Q0072 | 2024ApJS..273....3A | For simulated high-dynamic-range radio-interferometric imaging, how can images be reconstructed with higher SNR/logSNR while suppressing hallucinated structure? | U-Net | residual-to-residual DNN series | B | easy |
Q0073 | 2024ApJS..273....3A | For simulated radio-interferometric imaging, which learned reconstruction approach better preserves image-domain data fidelity while maintaining substantially higher reconstruction quality? | U-Net | plug-and-play learned-denoiser reconstruction | B | hard |
Q0074 | 2024MNRAS.531.1454K | When assigning optical emission-line equivalent widths to mock galaxy spectra, which empirical mapping is less tied to assumptions from a separate stellar-population model while giving comparable or slightly better EW predictions? | feed-forward neural network | gradient-boosted decision trees | A | hard |
Q0075 | 2024MNRAS.533.1976M | For classifying 1D optical spectra into quasars, galaxies, stars, and BAL quasars, which neural-network architecture gives higher classification accuracy when trained on the same spectra? | Transformer | 1D residual CNN | B | hard |
Q0076 | 2024A&A...687A.202W | For estimating wavefront aberrations from non-modulated pyramid wavefront-sensor images in adaptive optics, which neural-network architecture gives lower wavefront-estimation error? | global-context vision Transformer | ConvNeXt CNN | A | easy |
Q0077 | 2024A&A...687A.202W | For estimating wavefront aberrations from non-modulated pyramid wavefront-sensor images in adaptive optics, which neural-network architecture gives lower wavefront-estimation error? | wavefront-sensing CNN | global-context vision Transformer | B | easy |
Q0078 | 2024MNRAS.528.5890P | For estimating stellar surface gravity from stellar light curves, which neural-network architecture gives lower prediction error? | convolution-augmented Transformer | ResNet CNN | A | easy |
Q0079 | 2024MNRAS.528.5890P | For estimating stellar surface gravity from stellar light curves, which neural-network architecture gives lower prediction error? | convolution-augmented Transformer | convolution-only network | A | easy |
Q0080 | 2024MNRAS.528.5890P | For estimating stellar surface gravity from stellar light curves, which neural-network architecture gives lower prediction error? | convolution-augmented Transformer | self-attention-only Transformer | A | easy |
Q0081 | 2024MNRAS.528.5890P | For estimating stellar surface gravity from stellar light curves, which neural-network architecture gives lower prediction error? | convolution-only network | self-attention-only Transformer | B | easy |
Q0082 | 2024MNRAS.528.5890P | For estimating stellar surface gravity across the red-giant branch from stellar light-curve information, which machine-learning approach gives lower RMSE and fewer outlier predictions? | Transformer | k-nearest-neighbor regression | A | hard |
Q0083 | 2024AJ....167..118R | For compressing Solar System object feature vectors used in anomaly searches, which dimensionality-reduction method preserves the features more accurately? | principal component analysis | deep autoencoder | B | easy |
Q0084 | 2024AJ....167..118R | For compressing Solar System object feature vectors used in anomaly searches, which dimensionality-reduction method preserves the features more accurately? | non-negative matrix factorization | deep autoencoder | B | easy |
Q0085 | 2024ApJS..272...14H | Within a large survey variable-source classification system, for identifying RR Lyrae ab variable stars from light-curve features, which supervised ML method gives higher precision and recall? | XGBoost gradient-boosted decision trees | convolutional/dense neural network | A | hard |
Q0086 | 2024ApJS..272...14H | Within a large survey variable-source classification system, for identifying Delta Scuti variable stars from light-curve features, which supervised ML method gives higher precision and recall? | XGBoost gradient-boosted decision trees | convolutional/dense neural network | A | hard |
Q0087 | 2024ApJS..272...14H | Within a large survey variable-source classification system, for identifying active galactic nuclei from light-curve features, which supervised ML method gives higher precision and recall? | convolutional/dense neural network | XGBoost gradient-boosted decision trees | B | hard |
Q0088 | 2024A&A...681A..68E | For identifying galaxy-galaxy strong-lensing systems in high-resolution space-like images when less prominent lensing features are included, which CNN architecture gives higher classification accuracy and AUC? | residual CNN | inception CNN | B | hard |
Q0089 | 2024A&A...692A.260R | For AGN photometric-redshift estimation in wide-area survey data where host-galaxy and nuclear emission can be spatially mixed, which ML approach gives lower scatter and outlier fraction on an independent blind sample? | image-based CNN ensemble | fully connected neural network using catalog multi-aperture photometry | A | hard |
Q0090 | 2024AJ....168..234L | For identifying true stellar flares from flare-candidate profile features, which supervised classifier is more reliable? | XGBoost gradient-boosted decision trees | fully connected deep neural network | A | easy |
Q0091 | 2024AJ....168..234L | Among tree-based classifiers for identifying true stellar flares from flare-candidate profile features, which method better suppresses false positives without losing flare recovery? | XGBoost gradient-boosted decision trees | Random Forest | A | easy |
Q0092 | 2024A&A...690A.326L | For identifying galaxy mergers in high-resolution galaxy images, which CNN-based consensus classifier gives a better purity-completeness balance on visually classified real-galaxy test sets? | dual-CNN consensus classifier with a single decision cutoff | dual-CNN consensus classifier with ambiguous-case rejection | B | hard |
Q0093 | 2024MNRAS.532.1902R | For mitigating imaging systematics in photometric galaxy-clustering measurements used to constrain local primordial non-Gaussianity, which data-driven regression treatment better passes residual-systematics tests? | linear multivariate regression | feedforward neural-network regression | B | hard |
Q0094 | 2024RAA....24d5025L | For identifying cloudy-contaminated frames in light-curve preprocessing, when a classifier is trained on cloudy-frame data from one observing night and evaluated on frames from another observing night, which supervised classifier achieves higher classification accuracy? | support-vector machine | LightGBM | A | hard |
Q0095 | 2024ApJ...967..152A | For predicting galaxy star formation rates from broad-band galaxy images, which ResNet50-based architecture achieves higher test-set prediction accuracy? | sequential ResNet50 CNN chain | standalone ResNet50 CNN | A | hard |
Q0096 | 2024MNRAS.527.1240H | For inferring plane-of-sky magnetic-field orientation from spectroscopic velocity-channel maps under comparable molecular-cloud turbulence conditions, which CNN architecture gives lower median angular uncertainty? | 2D convolution-pooling CNN | 3D residual U-Net autoencoder | A | hard |
Q0097 | 2024ApJ...964..130J | For photometric redshift estimation in cosmology, which machine-learning pipeline gives lower RMS error and fewer outlier predictions? | Random Forest | Bayesian neural network with uncertainty-based rejection | B | easy |
Q0098 | 2024ApJ...964..130J | For photometric redshift estimation in cosmology, which machine-learning pipeline gives lower RMS error and fewer outlier predictions? | standard neural network | Bayesian neural network with uncertainty-based rejection | B | easy |
Q0099 | 2024ApJ...964..130J | For photometric redshift estimation in cosmology, which machine-learning pipeline gives lower RMS error and fewer outlier predictions? | XGBoost | Bayesian neural network with uncertainty-based rejection | B | easy |
Q0100 | 2024A&A...686A.144B | For filtering radio single-pulse candidates in magnetar observations before applying source-specific RFI rejection, which neural-network classification strategy gives the higher F-score? | single neural-network classifier | three-network voting ensemble | B | easy |
End of preview. Expand in Data Studio
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.
- Downloads last month
- 12