Modotte

---

Note : This datset is the part of a lineup MathX by Modotte you can get a lots of datasets on this same linup main focus is to provide very high quality datasets for model training and finetuning

This dataset is curated from high-quality public sources and enhanced with synthetic data from both closed and open-source models. It serves as a strong foundation for instruction-based model tuning and fine-tuning, offering one of the most refined and extensive corpora available.

Important:

• Num of rows with Think: 2022916
• Num of rows with No Think: 18229906

Key Features

• Scale: 20 million examples of highly curated step-by-step thinking data
• Diversity: Comprehensive coverage of mathematical domains from basic arithmetic to advanced calculus
• Quality: Multi-stage filtering and verification processes
• Reasoning: Step-by-step solutions with detailed mathematical thinking
• Accuracy: RL-verified answers with correctness validation

Dataset Overview

MathX is a meticulously curated mathematical reasoning dataset designed specifically for instruction-based model tuning and fine-tuning of existing models with enhanced thinking capabilities. This represents the largest and most comprehensively filtered corpus of publicly available mathematical reasoning data.

How to use?

pip install -U datasets fsspec

from datasets import load_dataset
dataset = load_dataset("Modotte/MathX-20M")

Key Features

• Scale: 20 million examples of highly curated step-by-step thinking data
• Diversity: Comprehensive coverage of mathematical domains from basic arithmetic to advanced calculus
• Quality: Multi-stage filtering and verification processes
• Reasoning: Step-by-step solutions with detailed mathematical thinking
• Accuracy: RL-verified answers with correctness validation

Data Curation Process

This dataset has been carefully constructed through a multi-source approach:

Data Sources

• High-Quality Existing Datasets: Curated from multiple premium mathematical datasets available online(Nvidia, Openr1, Modotte)
• Synthetic Generation: Generated using both closed-source and open-source language models(Modotte)
• Expert Validation: Human-verified mathematical solutions and explanations(Modotte)

Filtering Pipeline

Our rigorous filtering process includes:

1. Deduplication: Removal of duplicate problems and solutions
2. Normalization: Lowercasing and text standardization
3. Stopword Processing: Intelligent removal of non-essential words
4. Quality Scoring: Multi-dimensional quality assessment
5. Answer Verification: Reinforcement Learning-based answer validation
6. Content Filtering: Removal of inappropriate or incorrect content

Problem Complexity Distribution

• Basic Level (30%): Fundamental mathematical concepts and operations
• Intermediate Level (30%): Multi-step problems requiring reasoning chains
• Advanced Level (40%): Complex mathematical challenges and proofs

Mathematical Domains Covered

• Arithmetic and Number Theory, Algebra and Polynomial Mathematics, eometry and Trigonometry, Calculus and Analysis

Note : domains are for reference only the actual data is very diverse and covers more domains than stated actual data have more complex and high level questions than stated.

Use cases

• Fine-tuning mathematical reasoning capabilities in language models
• Training instruction-following models with mathematical focus
• Benchmarking model performance on mathematical reasoning tasks
• Research in mathematical AI and automated theorem proving
• Educational applications requiring step-by-step mathematical explanations

Dataset Format

Each example contains:

• Problem Statement: Clear mathematical question or challenge
• Step-by-Step Solution: Detailed reasoning process
• Final Answer: Verified correct solution

Quality Assurance

• RL Verification: All answers verified using reinforcement learning techniques
• Correctness Guarantee: Only problems with verified correct answers are included
• Human Review: Sample validation by mathematical experts
• Automated Checks: Computational verification where applicable

Performance Metrics

Models trained on this dataset show significant improvements in:

• Mathematical reasoning accuracy
• Step-by-step explanation quality
• Problem-solving methodology
• Cross-domain mathematical transfer

Acknowledgments

Special thanks to our partners and contributors:

• NVIDIA - Reference datsets and also the MathX contains many examples taken from Nvidia's existing datasets
• Openr1 - Reference datsets and also the MathX contains many examples taken from Openr1's existing datasets
• Modotte Team - Dataset curation and quality assurance alsong with some extra currated examples

Citation

Anyone can freely use and modify this dataset

License

This dataset is released under [ MIT - Lisence ].

@dataset{mathx2024,
  title={MathX: Large-Scale Mathematical Reasoning Dataset},
  author={Parvesh Rawal at Modotte},
  year={2024},
  publisher={Modotte},
  url={https://huggingface.co/datasets/Modotte/MathX}
}

Contact

For questions, suggestions, or collaboration opportunities:

• Email: Modotte
• Twitter: [@Modotte]
• GitHub: [Modotte]

---

Built with ❤️ by Modotte - Advancing AI through high-quality data