--- dataset_info: features: - name: id dtype: int64 - name: uid dtype: string - name: question dtype: string - name: permutation_idx dtype: int64 - name: choices list: string - name: labels list: int64 - name: prompt dtype: string - name: expected_output dtype: string splits: - name: train num_bytes: 399645 num_examples: 579 download_size: 103527 dataset_size: 399645 configs: - config_name: default data_files: - split: train path: data/train-* license: cc-by-4.0 task_categories: - question-answering - text-generation language: - en tags: - astronautics - space - engineering - multiple-choices - astrodynamics pretty_name: astro-mcq size_categories: - n<1K --- # Astro-MCQ Dataset **Astro-MCQ** is the first dataset in the upcoming **AstroBench** collection, a suite of domain-specific benchmark datasets for evaluating small and large language models (SLMs and LLMs) in space mission engineering and astronautics. ## Overview Astro-MCQ is a multiple-choice question dataset designed to evaluate language model performance across key topics in astronautics, including: - Orbital mechanics - Space propulsion - Space environment and its effects - Spacecraft systems and design - Mission operations - Human spaceflight - Launchers and more. `astro-mcq` is a revision of the following previous dataset https://huggingface.co/datasets/patrickfleith/Astro-mcqa/blob/main/README.md ## Purpose This dataset enables comparative assessment of LLM capabilities in the space engineering domain. It helps application developers and researchers answer critical questions: - **Model selection**: Which LLM performs best for your specific astronautics subdomain? - **Study configuration optimization**: What model size, quantization level, and prompting strategy work best? - **Capability assessment**: How do different small-to-medium open-weight models compare in reasoning and domain knowledge? - **Fine-tuning evaluation**: How effective is domain adaptation or specialized fine-tuning? ## Use Cases ### Recommended Uses - **Model evaluation and benchmarking**: Compare performance across different LLMs - **Quantization testing**: Assess how compression affects domain-specific performance - **Prompt engineering**: Test and optimize prompting strategies - **Domain adaptation**: Evaluate effectiveness of fine-tuning approaches - **Model auditing**: Verify model capabilities before deployment ### Evaluation Methods The dataset supports two evaluation approaches: - **Loglikelihood-based evaluation**: Token probability scoring - **Generative evaluation**: Free-form response assessment (e.g., model-as-a-judge) ### Not Recommended For - **Training or fine-tuning**: Dataset size is too limited for effective model training - **Sole training resource**: Could potentially be combined with other datasets for meta-learning, but not suitable as primary training data ## Quick Start ### Explore Online Browse the dataset: https://huggingface.co/datasets/patrickfleith/astro-mcq/viewer/default/train ### Download and Use **Manual download:** https://huggingface.co/datasets/patrickfleith/astro-mcq **Python:** ```python from datasets import load_dataset dataset = load_dataset("patrickfleith/astro-mcq") ``` ## About AstroBench Astro-MCQ is the first release in the AstroBench collection. AstroBench aims to provide comprehensive evaluation tools for assessing language models in space mission engineering contexts, covering multiple task types and difficulty levels tailored to real-world astronautics applications. ### Structure The dataset contains **193 expert-created multiple-choice questions**, with **3 permutations per question** (randomized choice order), resulting in approximately **579 evaluation instances**. This permutation strategy helps mitigate the sensitivity of evals to model bias position (i.e. their tendency to prefer certain propositions) Each instance includes the following fields: - **uid**: Unique identifier string for the original question (UUID format), shared across permutations of the same question - **question**: The question text as a string - **permutation_idx**: Integer (0-2) indicating which permutation of the question this is - **choices**: List of answer choices as strings. Questions can have multiple correct answers. - **labels**: List of integers (0 or 1) corresponding to each choice. 0 = incorrect, 1 = correct. Multiple labels can be 1. - **prompt**: Pre-formatted prompt string ready for LLM evaluation (includes question and formatted choices) - **expected_output**: The correct answer(s) formatted as expected (e.g., "A. Choice text" or "A. Choice\nB. Choice" for multiple correct answers) ### Languages All instances in the dataset are in english   ### Size - **193** expert-created unique questions or retrieved from master-level courses in space mission design and operations. - **~579** total evaluation instances (3 permutations per question) - Questions filtered from the original dataset to guarantee at least **3 choices** per question and at least **1 correct answer**   ### Question Types - **Knowledge-based**: Questions testing domain knowledge, facts,, in space science and engineering - **Reasoning**: Questions requiring logical reasoning and problem-solving, and understanding of physics.. - **Computational**: Questions requiring mathematical operations with numerical results (exam-style) - **Multiple-answer**: Some questions have multiple correct choices (multi-select format) ### Topics Covered Comprehensive coverage across space engineering subdomains: - Orbital mechanics and trajectories - Space propulsion systems - Mission operations and design - Human spaceflight - Space environment and effects - Spacecraft systems and subsystems - Communication and link analysis - Space project lifecycle - Launch systems and more # USAGE AND GUIDELINES #### License AstroMCQ © 2025 by Patrick Fleith is licensed under Creative Commons Attribution 4.0 International #### Restrictions No restriction. Please provide the correct attribution following the license terms.   #### Citation P. Fleith, "Astro-MCQ: A Multiple-Choice Question Benchmark Dataset for Evaluating LLMs in Space Mission Engineering and Astronautics," (2025).   #### Update Frequency May be updated based on feedbacks. If you want to become a contributor, let me know.   #### Have a feedback or spot an error? Use the community discussion tab directly on the huggingface `astro-mcq` dataset page.   #### Contact Information Reach me here on the community tab or on LinkedIn (Patrick Fleith) with a Note. #### Current Limitations and Future Work - Only 193 multiple choice questions and answers. This makes it useless for fine-tuning purposes although it could be integrated as part of a larger pool of datasets compiled specifically for fine-tuning. - While being a decent size enabling LLM evaluation, space engineering expert time is scarce and expensive. On average it takes 8 minutes to create one MCQA example. Having more examples would be much better for robustness. - The dataset might be biased toward the very low number of annotators. - The dataset might be biased toward European Space Programs. - The dataset might not cover all subsystems or subdomains of astronautics, although we tried to do our best to cover the annotators' domains of expertise. - No peer review. Ideally, we would like to have a Quality Control process to ensure high quality and correctness of each example in the dataset. Given the limited resources, this is not yet possible. Feel free to come and contribute if you feel this is an issue.