Math Problems · AI-MO
NuminaMath-TIR
A dataset of mathematical problems paired with tool-integrated reasoning paths and code execution outputs.
What NuminaMath-TIR is
NuminaMath-TIR is a curated mathematical reasoning dataset created by Project Numina (AI-MO). Designed for mathematical problem-solving tasks, it focuses on tool-integrated reasoning (TIR), an approach that embeds executable code within textual reasoning chains.
The dataset addresses the computational and cost challenges of generating high-grade annotations for tool use in mathematics.
For researchers developing language models for scientific reasoning and mathematics, this dataset provides verified solutions where computational steps are handled by programmatic execution rather than mental calculation alone.
How it was made
An automated pipeline driven by GPT-4 was employed to generate ToRA-style reasoning paths that interleave text with code.
During generation, the embedded code was executed in real time, feeding outputs back into the reasoning prompt until the final solution was reached. To maintain accuracy and consistency, solutions were verified against reference answers, filtering out any mismatched results across three iterative generation cycles.
Key facts
- Format
- Parquet
- Language
- English
- Task category
- Text generation
- Modality
- Text
- Release date
- 16 July 2024
What you can use it for
- Tool-Integrated Model Fine-Tuning
- Training language models to generate executable code blocks and interpret runtime output during mathematical problem solving.
- Mathematical Reasoning Research
- Investigating how interleaving natural language reasoning with programmatic tools affects numerical accuracy in complex problem domains.
- Competition Mathematics Benchmarking
- Evaluating algorithmic reasoning systems on problems drawn from mathematics competitions requiring numerical and integer answers.
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-TIR at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 72,441 rows of default/train
messages
list- min
- 2
- median
- 2
- mean
- 2
- max
- 2
problem
string_text- min
- 11
- median
- 179
- mean
- 217.83
- max
- 4,309
solution
string_text- min
- 146
- median
- 1,686
- mean
- 2019.01
- max
- 12,279
Strengths and limits
Good at
- All reasoning paths are verified by executing the generated code and confirming that final answers match reference solutions.
- Data quality is supported by repeating the generation and filtering process three times.
- Focuses specifically on problems with numerical outcomes, where tool execution directly aids correctness.
Watch out for
- Restricted primarily to problems with numerical, largely integer, target answers.
- Reasoning paths depend on synthetic trajectories generated by GPT-4 rather than human-written explanations.
Questions
Can I look inside NuminaMath-TIR 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-TIR under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
What does tool-integrated reasoning mean in this dataset?
Tool-integrated reasoning refers to reasoning paths that write and execute computer code, incorporating the computational outputs back into the step-by-step solution.
How are the problems sourced?
The problems are selected from the NuminaMath-CoT dataset, focusing specifically on entries that require numerical answers.
How was accuracy validated during data creation?
Generated solutions were executed and strictly filtered out if the final answer did not match the known ground truth, repeating this verification procedure three times.
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, and Stanislas Polu. NuminaMath TIR. Hugging Face repository, 2024.
Written from NuminaMath-TIR'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.