metadata

license: apache-2.0
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - math
  - post-training
  - RL
  - verifiable
  - reasoning
pretty_name: NuminaMath 1.5 RL Verifiable
dataset_info:
  features:
    - name: problem
      dtype: string
    - name: solution
      dtype: string
    - name: answer
      dtype: string
    - name: problem_type
      dtype: string
    - name: question_type
      dtype: string
    - name: problem_is_valid
      dtype: string
    - name: solution_is_valid
      dtype: string
    - name: source
      dtype: string
    - name: synthetic
      dtype: bool
  splits:
    - name: train
      num_bytes: 188626432
      num_examples: 131063
  download_size: 85718743
  dataset_size: 188626432
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Card for NuminaMath-1.5-RL-Verifiable

Dataset Description

Homepage: https://huggingface.co/datasets/nlile/NuminaMath-1.5-RL-Verifiable
Repository: NuminaMath-1.5-RL-Verifiable
Based on: NuminaMath-1.5

Dataset Summary

NuminaMath-1.5-RL-Verifiable is a curated subset of the NuminaMath-1.5 dataset, specifically filtered to support reinforcement learning applications requiring verifiable outcomes. This collection consists of 131,063 math word problems from the original dataset that meet strict filtering criteria: all problems have definitive numerical answers, validated problem statements and solutions, and come from high-quality, non-synthetic sources.

The filtering process removes multiple-choice questions, proofs, problems without clear numerical answers, and all synthetic content, while preserving the rich diversity of mathematical domains from the original collection.

Filtering Methodology

The dataset was created by applying the following filters to the original NuminaMath-1.5 dataset:

Removed question types: Multiple-choice questions and proofs
Answer validation: Retained only problems with non-empty, numerical answers (excluded 'proof', 'notfound' answers)
Source selection: Excluded potentially lower-quality sources (cn_k12, orca_math, synthetic_math, metamath)
Quality filters: Retained only problems with validated problem statements and solutions
Authenticity: Excluded all synthetic problems

These filtering steps reduced the original dataset from 896,215 problems to 131,063 problems (approximately 14.6% of the original dataset), all with verifiable outcomes.

Dataset Structure

Data Instances

Each instance in the dataset contains:

A math word problem statement
A Chain of Thought (CoT) solution
A definitive numerical answer
Problem metadata including math domain type

Data Fields

problem: Text description of the mathematical problem
ref_solution: Step-by-step Chain of Thought (CoT) solution
answer: Definitive numerical result
problem_type: Mathematical domain (Algebra, Geometry, Number Theory, etc.)
question_type: Always "math-word-problem" in this filtered dataset
source: Origin of the problem (olympiads, cn_contest, aops_forum, etc.)
problem_is_valid: Always "Yes" in this filtered dataset
solution_is_valid: Always "Yes" in this filtered dataset
synthetic: Always false in this filtered dataset

Dataset Statistics

Distribution by Source

Source	Problem Count
olympiads	92,487
cn_contest	15,828
aops_forum	15,092
amc_aime	4,893
inequalities	1,145
olympiads_ref	1,001
number_theory	617
Total	131,063

Distribution by Problem Type

Problem Type	Problem Count	Percentage
Algebra	42,972	32.79%
Geometry	31,405	23.96%
Number Theory	22,071	16.84%
Combinatorics	17,144	13.08%
Logic and Puzzles	7,250	5.53%
Calculus	4,954	3.78%
Inequalities	4,000	3.05%
Other	1,267	0.97%

Detailed Breakdown by Problem Type and Source

Click to expand detailed breakdown

Algebra

olympiads: 31,752
cn_contest: 6,776
amc_aime: 1,886
aops_forum: 1,684
inequalities: 531
olympiads_ref: 265
number_theory: 78

Geometry

olympiads: 22,091
cn_contest: 4,377
aops_forum: 3,316
amc_aime: 1,454
olympiads_ref: 99
inequalities: 60
number_theory: 8

Number Theory

olympiads: 14,848
aops_forum: 3,614
cn_contest: 1,916
amc_aime: 744
number_theory: 489
olympiads_ref: 329
inequalities: 131

Combinatorics

olympiads: 11,219
aops_forum: 3,176
cn_contest: 1,724
amc_aime: 612
olympiads_ref: 266
inequalities: 125
number_theory: 22

Logic and Puzzles

olympiads: 5,677
aops_forum: 1,197
cn_contest: 212
amc_aime: 136
inequalities: 16
number_theory: 7
olympiads_ref: 5

Calculus

olympiads: 3,894
aops_forum: 907
cn_contest: 139
inequalities: 8
amc_aime: 4
olympiads_ref: 1
number_theory: 1

Inequalities

olympiads: 2,292
aops_forum: 717
cn_contest: 657
inequalities: 273
olympiads_ref: 34
amc_aime: 25
number_theory: 2

Other

olympiads: 714
aops_forum: 481
amc_aime: 32
cn_contest: 27
number_theory: 10
olympiads_ref: 2
inequalities: 1

Original NuminaMath-1.5 Source Breakdown

source	problems	question_type:proof	question_type:mcq	question_type:word
olympiads	197084	62970	13529	117845
olympiads_ref	3638	2246	nan	1392
amc_aime	5872	208	4374	963
aops_forum	67841	24532	5924	33486
cn_contest	29944	8663	5602	15649
inequalities	7314	5780	49	1478
number_theory	4043	2591	15	1239
cn_k12	268819	3966	115800	149010
orca_math	151934	1	17	151916
synthetic_math	148712	41	1057	147612
metamath	11014	nan	82	10932
Total	896215	110998	146449	631522

Additional Information

Licensing Information

The dataset follows the licensing of the original NuminaMath-1.5 dataset and is available under the Apache License, Version 2.0.

Citation Information

@misc{nlile2025numinamath15rlverifiable,
  author = {nlile},
  title = {NuminaMath-1.5-RL-Verifiable},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Dataset Repository},
  howpublished = {\url{https://huggingface.co/datasets/nlile/NuminaMath-1.5-RL-Verifiable}}
}

@misc{numina_math_datasets,
  author = {Jia LI and Edward Beeching and Lewis Tunstall and Ben Lipkin and Roman Soletskyi and Shengyi Costa Huang and Kashif Rasul and Longhui Yu and Albert Jiang and Ziju Shen and Zihan Qin and Bin Dong and Li Zhou and Yann Fleureau and Guillaume Lample and Stanislas Polu},
  title = {NuminaMath},
  year = {2024},
  publisher = {Numina},
  journal = {Hugging Face repository},
  howpublished = {\url{https://github.com/project-numina/aimo-progress-prize/blob/main/report/numina_dataset.pdf}}
}