MathDial

MathDial is a research dataset from the Language, Reasoning and Education lab at ETH Zurich, released with the Findings of EMNLP 2023 paper by Jakub Macina, Nico Daheim and colleagues. It collects about 2,861 one-to-one tutoring dialogues (roughly 2,260 training and 599 test), each grounded in a multi-step maths word problem from GSM8K and a specific student confusion. The conversations were built by pairing 91 human teachers, who role-played the tutor, with a language model (InstructGPT, text-davinci-003) prompted to act as a student holding a realistic misconception. Every tutor turn is labelled with a pedagogical move from a four-part taxonomy: focus, probing, telling and generic, so the data records not just what was said but the teaching strategy behind it. You access it freely from Hugging Face (CSV and JSONL in the repository, auto-converted Parquet on the Hub, loadable with the datasets library) or from the GitHub repository. For RAG it is a small, clean, well-structured corpus of annotated pedagogy rather than subject content. For an AI tutor its value is specific: where most maths sources here supply material to teach from, MathDial models Socratic remediation, showing how an expert guides a learner from a wrong answer to the right one without simply revealing it, which is directly useful for training tutoring behaviour, ranking candidate responses and evaluating whether a model tutors or just solves. On licensing there is a discrepancy worth noting: the Hugging Face dataset card tags it CC BY 4.0, while the GitHub repository states CC BY-SA 4.0. Either way attribution is required and commercial use is permitted, but you should verify whether a share-alike obligation applies before building on it commercially. Because the student turns are model-generated rather than real learners, it carries no real-pupil privacy constraint, unlike bridge-math-remediation and case-network. It complements the curriculum text in openstax, ck-12, khan-academy and siyavula and the problem sets in amps, megamath, stackmathqa and naturalproofs, with common-core and ngss providing the standards its problems map to.