ScienceQA Multimodal Science Questions

## Overview ScienceQA is a multimodal multiple-choice question answering benchmark covering natural science, social science, and language science topics from elementary and high school curricula. It contains ~21,000 examples with questions, answer choices, optional images, contextual lectures, and chain-of-thought solution explanations. The dataset is distributed in Parquet format with image and text modalities and is split into train/validation/test. All content is in English. ## Schema - `question` — string — the science question text - `choices` — list[string] — multiple-choice answer options - `answer` — int — index of the correct choice - `image` — Image — optional accompanying image (many examples are text-only) - `hint` — string — optional context or hint - `task` — string — task type (closed/open domain) - `grade` — string — grade level (e.g., grade1–grade12) - `subject` — string — top-level subject (natural science, social science, language science) - `topic` — string — fine-grained topic - `category` — string — sub-category within topic - `skill` — string — specific skill tested - `lecture` — string — background lecture text - `solution` — string — chain-of-thought explanation of the correct answer ## Sources - HuggingFace: https://huggingface.co/datasets/derek-thomas/ScienceQA — license: CC-BY-SA-4.0 - Original paper: Lu et al., "Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering" (NeurIPS 2022), arXiv:2209.09513 ## Methodology Questions were sourced from elementary and high school science curricula and annotated by domain experts with answer choices, lectures providing background context, and chain-of-thought solutions explaining the reasoning. The derek-thomas mirror repackages the original ScienceQA release into Parquet with embedded images for easier streaming via the `datasets` library. ## Known gaps & limitations English-only; coverage is biased toward US K-12 science curriculum topics and may not generalize to other education systems. Image quality and presence varies — a substantial fraction of examples are text-only. The dataset is widely used as a public eval benchmark, so leakage into pretraining corpora of recent LLMs/VLMs is likely; buyers training on this should not also evaluate on it without contamination checks. ## Intended use & out-of-scope - Intended: multimodal VLM evaluation, chain-of-thought fine-tuning, science QA research, instruction tuning with explanations. - Out-of-scope: not deduplicated against common eval suites — leakage risk for benchmark training; not suitable as a sole signal for general science knowledge given K-12 scope. _Federated dataset: 3 parquet shards, 597.1 MB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: ScienceQA: Multimodal Science Question Answering with Chain-of-Thought 21K multimodal multiple-choice science questions with images, lectures, and chain-of-thought explanations spanning natural science, social science, and language science. Widely used for VLM evaluation and CoT fine-tuning.