Math Problems · AI-MO
NuminaMath-CoT
NuminaMath with Chain-of-Thought reasoning annotations.
What NuminaMath-CoT is
Each problem is paired with a solution structured in a Chain-of-Thought (CoT) format, making explicit the logical progression required to reach the final answer.
The dataset spans diverse mathematical sources, including Chinese high-school exercises, online discussion forums, and competitive problems from United States and international mathematics olympiads. It addresses the shortage of large, diverse, and uniformly structured reasoning datasets needed to teach language models rigorous mathematical problem-solving.
How it was made
The dataset was built from online exam paper PDFs and mathematical discussion forums. Raw inputs underwent optical character recognition (OCR) where necessary, followed by segmentation into discrete problem-solution pairs.
Non-English problem pairs were translated into English. The solutions were subsequently realigned into a standardised Chain-of-Thought format, and their final answers were uniformly formatted. The resulting collection contains 859,608 samples stored in Parquet format, covering both natural forum and competition problems as well as synthetic expansions.
Key facts
- Total samples
- 859,608
- Modality
- Text
- Language
- English
- Format
- Parquet
- Primary task
- Text generation / mathematical reasoning
What you can use it for
- Supervised fine-tuning for mathematical reasoning
- Fine-tune generative language models on structured, multi-step chain-of-thought solutions to improve intermediate deduction and final answer accuracy.
- Competition mathematics benchmarking and instruction
- Train and evaluate models on advanced problems drawn from AMC, AIME, and international olympiad competitions.
- Automated reasoning pipeline development
- Use standard step-by-step rationales to train systems that segment, check, and produce logical proofs and calculations.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from AI-MO/NuminaMath-CoT at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 859,494 rows of default/train
messages
list- min
- 2
- median
- 2
- mean
- 2
- max
- 2
problem
string_text- min
- 10
- median
- 213
- mean
- 251.58
- max
- 6,097
solution
string_text- min
- 13
- median
- 996
- mean
- 1170.84
- max
- 10,391
source
string_label- cn_k1232%
- synthetic_math20%
- orca_math18%
Strengths and limits
Good at
- Contains 859,608 processed problems across elementary to olympiad-level mathematics.
- Provides uniform Chain-of-Thought reasoning steps and standardised final answer formatting for every problem.
- Brings together diverse sources including Chinese high-school curricula, forum discussions, established benchmarks (GSM8K, MATH), and synthetic problems.
Watch out for
- A portion of the data relies on OCR from original PDFs, which can introduce transcription artefacts if not fully resolved.
- Translations into English from Chinese sources may introduce linguistic or notational discrepancies relative to standard Western curricula.
Questions
Can I look inside NuminaMath-CoT without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is NuminaMath-CoT under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What mathematical levels are included in NuminaMath-CoT?
The dataset spans secondary school exercises up to advanced international olympiad competitions, including sources such as Chinese K-12, AMC, AIME, and AoPS forums.
What format are the solutions provided in?
All solutions are processed into an English Chain-of-Thought (CoT) format with standardised final answer formatting.
Are synthetic data included in the dataset?
Yes. The dataset includes 62,111 synthetic AMC problems and 167,895 synthetic MATH problems alongside human-curated competition and textbook problems.
How to cite it
Jia LI, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Costa Huang, Kashif Rasul, Longhui Yu, Albert Jiang, Ziju Shen, Zihan Qin, Bin Dong, Li Zhou, Yann Fleureau, Guillaume Lample, Stanislas Polu. NuminaMath. 2024.
Written from NuminaMath-CoT's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.