Math Problems · meta-math
MetaMathQA
A mathematical question-answering dataset augmented from GSM8K and MATH to train and evaluate mathematical reasoning.
What MetaMathQA is
MetaMathQA is a text dataset designed for training and evaluating mathematical reasoning in large language models. Published by meta-math, it targets the challenge of teaching computational models step-by-step reasoning on complex quantitative problems.
The corpus is built by augmenting questions and answers strictly from the training partitions of two standard benchmarks: GSM8K, which focuses on grade-school mathematics, and MATH, which comprises harder competition-level problems. No data is derived from the testing splits of either benchmark.
Models trained on MetaMathQA, such as MetaMath-7B, MetaMath-13B, and MetaMath-Mistral-7B, demonstrate marked performance gains when evaluated on standard mathematical problem-solving benchmarks like GSM8K Pass@1 and MATH Pass@1.
How it was made
MetaMathQA consists of question-answer pairs generated through data augmentation applied to the training sets of GSM8K and MATH. Every item contains an original question field that links the augmented instance back to its source problem in the respective training split.
It is intended for instruction fine-tuning using a prompt structure where the query is presented as an instruction followed by a step-by-step reasoning response prefix.
Key facts
- Modality
- Text
- Format
- JSON
- Source datasets
- GSM8K training set and MATH training set
- Primary paper
- arXiv:2309.12284
What you can use it for
- Instruction fine-tuning for mathematical reasoning
- Fine-tuning base language models such as Mistral-7B and LLaMA variants to generate multi-step mathematical solutions.
- Dataset combination for multi-task training
- Merging MetaMathQA with other instruction corpora, such as MathInstruct, to produce hybrid mathematical reasoning models.
- Mathematical QA benchmarking
- Evaluating the mathematical question-answering accuracy and chain-of-thought capabilities of trained language models.
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 meta-math/MetaMathQA at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 395,000 rows of default/train
original_question
string_text- min
- 16
- median
- 181
- mean
- 196.73
- max
- 4,309
query
string_text- min
- 0
- median
- 198
- mean
- 213.54
- max
- 4,409
response
string_text- min
- 15
- median
- 426
- mean
- 498.24
- max
- 5,367
type
string_label- GSM_Rephrased20%
- GSM_AnsAug20%
- MATH_AnsAug19%
Strengths and limits
Good at
- Augmented strictly from training splits, preventing benchmark contamination against GSM8K and MATH test sets.
- Includes an explicit original question field to trace each augmented item back to its training source.
- Consistently improves benchmark performance on GSM8K and MATH across various base model families, including LLaMA and Mistral.
Watch out for
- Problem topics and scope are constrained by the distributions of the underlying GSM8K and MATH source datasets.
- Optimal fine-tuning of models like Mistral-7B requires specific hyperparameter modifications, such as reducing the learning rate relative to LLaMA models.
Questions
Can I look inside MetaMathQA 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 MetaMathQA under?
Its authors declare mit. Read the terms on the original page before reusing the data.
Does MetaMathQA include any test set problems from GSM8K or MATH?
No. All examples in MetaMathQA are augmented solely from the training sets of GSM8K and MATH, ensuring test evaluations remain uncontaminated.
How can users trace the source of an augmented question?
Each data record contains an original question attribute that identifies the exact problem from which it was derived.
What prompt structure is recommended for training on this dataset?
The recommended format frames the problem within a task instruction template ending with the prompt prefix asking the model to think step by step.
Which language models have been trained with MetaMathQA?
Documented models fine-tuned on the dataset include MetaMath-7B, MetaMath-13B, MetaMath-Mistral-7B, Arithmo-Mistral-7B, OpenChat-3.5, CausalLM, Zephyr, and Ziya2.
How to cite it
Yu, Longhui, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 'MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models.' arXiv preprint arXiv:2309.12284 (2023).
Written from MetaMathQA'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.