OpenThoughts3-1.2M

OpenThoughts3-1.2M is an open reasoning dataset released in June 2025 by the Open Thoughts team, a collaboration of academic and industry researchers. It is the training corpus behind OpenThinker3-7B, which on release beat other similarly sized open-data reasoning models. The dataset holds 1.2 million rows: roughly 850,000 maths questions, 250,000 code questions and 100,000 science questions, each paired with a detailed chain-of-thought answer.

A key point is provenance. The questions were sourced and selected through more than a thousand ablation experiments documented in the accompanying paper, and the answers were generated by QwQ-32B, a reasoning model distributed under Apache 2.0. Because the answers are synthetic outputs of an openly licensed model rather than scraped human text, there is no hidden content licence or model-output restriction to trip over downstream.

Access is simple. It is published on Hugging Face in Parquet, so you can stream it with the datasets library, filter to a single domain, or pull the whole set (around 60 GB) for local training. There is no API beyond the standard Hub tooling, and no credentialing.

The obvious use is supervised fine-tuning of a reasoning model, teaching it to produce step-by-step traces for maths, code and science. For RAG it is less a knowledge base than a source of worked examples: you can index the question-and-solution pairs to ground a tutor or to seed few-shot prompts for a maths or coding assistant.

The licence is Apache 2.0 with no territory restriction, which permits commercial use and requires only that you preserve the licence and attribution notices. There is no share-alike obligation. Against the maths sources we already catalogue, it differs in kind: AMPS and StackMathQA offer human-written questions and solutions, while this set offers model-generated reasoning traces at scale, making it better suited to distilling reasoning behaviour than to factual retrieval.