ScienceQA
ScienceQA is a benchmark of 21,208 multimodal multiple-choice science questions drawn from elementary and high school curricula, released with the NeurIPS 2022 paper "Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering" (Pan Lu and colleagues). It spans three subjects, natural science, language science and social science, organised into 26 topics, 127 categories and 379 skills, so questions carry both a grade band (grades 1 to 12) and a fine-grained skill label. Crucially for teaching use, 83.9% of questions come with a grounded lecture and 90.5% with a detailed explanation, giving you not just the answer but the reasoning behind it. Many items include an image, a text passage or both.
Access is straightforward and free. The text (problems.json) lives in the GitHub repository, the images are distributed via Google Drive, and the dataset is also mirrored on Hugging Face Datasets and OpenDataLab. That makes it easy to pull the whole set for a build rather than scraping it yourself.
For RAG, each question plus its lecture and explanation is a compact, self-contained unit that maps cleanly onto retrieval, which is why we mark it chunked. For an AI tutor the value is the structure: the grade band supports curriculum sequencing and level-graded delivery, the 379 skills support mastery modelling and targeted practice, and the paired lectures and explanations provide the worked, step-by-step pedagogy a tutor needs to teach rather than merely grade. The multimodal items also suit tutors that reason over diagrams.
Note the licence carefully. The data is released under CC BY-NC-SA 4.0 (the repository ships a LICENSE-DATA file; the code is separately MIT). This is non-commercial and share-alike, so it cannot be used in a paid or revenue-generating tutor, attribution is required, and any derivative dataset you publish must carry the same licence. Where we list broadly reusable teaching material such as OpenStax, CK-12 or Khan Academy, or maths-focused sets like AMPS and StackMathQA, ScienceQA is the narrower, explanation-rich, level-tagged science complement, best kept to research and non-commercial projects.
Related sources
Achievement Standards Network
Machine-readable curriculum standards from US states, national bodies and other jurisdictions, modelled as an RDF graph of URI-addressable learning objectives with cross-jurisdiction alignments. Now run by D2L and free to use.
AGIEval
8,062 questions drawn from 20 official standardised exams (SAT, LSAT, GMAT, GRE, Gaokao, AMC/AIME and more) in English and Chinese, packaged as a human-centric benchmark for evaluating foundation models.
CASE Network (1EdTech)
A public registry of machine-readable learning-standard frameworks from all 50 US states and other issuing agencies, run by 1EdTech in the CASE JSON format. The digitally referenceable spine of what to teach, at which level and in what order, rather than the teaching content itself.
CEFR Companion Volume Descriptors
The Council of Europe's 2020 CEFR Companion Volume descriptor set, the de facto international standard for levelling language proficiency from Pre-A1 to C2. A spine of can-do statements for sequencing and grading language teaching.